CN103752534B - Intelligence feel digital image recognition sorting equipment and identification method for sorting - Google Patents

Abstract

The invention discloses a kind of intelligence feel digital image recognition sorting equipment, at least one group of LED is respectively arranged with above the position of measured piece mobile alignment, CCD camera is provided with above the axle center of LED, CCD camera is connected with PC, PC is connected with touch-screen, PC is connected with PLC, and the output of PLC connects with corresponding purge valve and (or) manipulator respectively by multiple magnetic valve, and each purge valve and (or) manipulator are arranged on delivery track side.The invention also discloses a kind of intelligence feel digital image recognition method for sorting.Device and method of the present invention, can ensure identification and the sorting of 100%, significantly improve operating efficiency and quality.

Description

Intelligence feel digital image recognition sorting equipment and identification method for sorting

Technical field

The invention belongs to Intelligent Recognition detection technique field, relate to a kind of intelligence feel digital image recognition sorting equipment, the invention still further relates to a kind of intelligence feel digital image recognition method for sorting.

Background technology

Machine vision technique have untouchable, real-time, speed are high, precision is high, strong interference immunity, can significantly reduce costs, improve speed of production and efficiency, have broad application prospects in Industrial Inferential Measurements and production control.But existing Machine Vision Detection identification is have preferably effect in the quiescent state substantially, in many situations, it is undetected that Detection results does not reach zero misjudgement zero, comprises following aspect:

1, the method realizing moving object segmentation mainly comprises: 1) background subtraction: completely can be partitioned into moving image rapidly, its deficiency is subject to light variable effect, and the renewal of background is crucial, is not suitable for the situation of cam movement; 2) optical flow method: the image that can detect self-movement, can be used for the situation of cam movement, but calculation of complex is consuming time, more difficultly realizes Real-Time Monitoring; 3) frame difference method: less by light variable effect, simple and quick, but complete Moving Objects can not be partitioned into, need to use Target Segmentation algorithm further.

2, article are at a distance of the detection of dynamic recognition methods of certain intervals.The vehicular traffic license plate identification of such as extensive use is under vehicle keeps certain safe spacing to move or stops static state, under specific color background, takes out license plate frame figure, then the character information of all number-plate numbers is implanted in Car license recognition software in advance, belong to the identification of template character match pattern, its false determination ratio is at 3%-10%.The model of prior art is all have very strong recognition capability in the ideal case, easily realize 1 accurate detection and Identification to 1 template and single individuality, and in actual applications, due to real time environment complicated situation, the interference of a variety of causes can be subject to, often there is matching operation failure, even if or provide erroneous matching separately seek his way, the founding mathematical models such as wherein Self Adaptive Control, fuzzy logic, neutral net carry out template matches obtain certain specific background under application.Due to picture processing speed 280-150ms/ place, software computing recognition speed 0.64-0.58s/, calculation of complex, length consuming time, efficiency is low, cost is high, can not adapt to multiple occasion.

The rapid image matching algorithm of current extensive use, such as images match, template matches, SSDA, projection matching etc. due to amount of calculation too large, long speed consuming time is slow especially, real-time is poor, and match time is oversize, cannot be applied in industrial production at a high speed.Have to adopt the higher high-precision hardware CCD camera of relative cost to cover the shortage, software recognition detection and the effect of hardware independent, often in practice, automatically can not identify the change of external items and environment, constantly allow manually to CCD camera, light, software recognition detection parametric synthesis adjustment can Results, number of repetition length how consuming time.

So existing software recognition detection continuous dynamic motion product, the ratio that undetected and false determination ratio maintenance is quite high, make it not implement and to apply.

3, the dynamic surface analyzing and processing of image.For image processing program, there is not dynamic image, the set of so-called dynamic image namely every frame image data, such as film and cartoon, 24-30 frame picture is flashed then concerning the eyes of people as long as each second is before eyes, it is exactly a complete continuous print animation, but concerning equipment, namely 24-30 frame picture have passed through projected area within 1 second, that is in fact image processing program carries out independent process to each frame pictorial information, just image belongs to transmission continuously, namely a frame connects a frame, concerning people just as animation, also be so-called dynamic process." degeneration " in dynamic imaging processes, refer to the impact due to imaging system various factors, the noise of digital picture is mainly derived from acquisition and the transmitting procedure of image, i.e. the noise jamming of transmission channel in the digitized process of Image Acquisition, the quality of imageing sensor and environmental condition, image transmitting process.Another situation is exactly motion blur, capital makes image recognition misjudge the generation of failing to judge, in order to solve such problem, adapt to multiple occasion carry out image dynamic surface analyzing and processing application, people wish filter function be dynamic, adjustable, can self-teaching.It is very difficult for designing such wave filter, does not also have the wave filter of the dynamic surface of image to produce up till now.

Also image recognition software and hardware are combined by nobody automatically, the intelligently guiding formula vision system of the exploitation even if try to be the first in current countries in the world, number of processes (the binaryzation that the same wave filter of Image semantic classification is maximum 9 times, difference, bright cloth just, contrast is changed, real time differential, real-time deep or light correction, Fuzzy Processing only 1 time), the maximum 13 layers of (binaryzation of pretreatment combination size, difference, spot only can use 1 time), the beat of intelligently guiding formula vision system Dynamic Recognition automated production, frequency only under the low speed effectively (still has undetected inevitable with false retrieval), for two of dynamic motion close products, or product surface has greasy dirt, the product that background can constantly change, cannot 100% positive and negative detecting product, the beat of high speed automated production, frequency cannot realize.

On current equipment rail or article continuous detection one by one on conveyer belt, rely on High-Speed Hardware shooting or photograph visual identifying system can realize consecutive objects dynamic motion in theory change static vision identification into and detect continuously, due to the costliness of its price and maintenance cost, and to applied environment and operation requirements harshness, there is no this respect application report up till now.

4, contact (two gold) realizes switch, overload long delay as miniature circuit breaker and moves two segment protect functions indispensable companion short circuit wink.The correct assembling of contact (two gold) and whether intact, directly and breaker quality and apparent size closely bound up, in contact (two gold) production operation, very important work is exactly the two golden flexure plane of how management and control, the identification of the positive and negative of contact operates with sorting, continue to use manual operation visual detection up till now always, repeatedly checking of dimensions is repeated to parts such as same contacts (two gold), mark, the characteristic informations such as identified surface, the product information having carried out each intermediate link of contact (two gold) detects, identify and sorting, there is operating personnel's eye fatigue, artificial uncertain factor such as changeable in mood grade, need the manpower of at substantial, and the probability of makeing mistakes is higher, rework cost is very high, scrappage is high.

And the detection modes such as the detection of mechanical touch formula, optoelectronic switch detection, because installation and debugging are difficult, accuracy of detection is not high, detection speed slow and it is unstable to detect, be difficult to realize testing product accuracy, so, do not substitute the automatic recognition detection Sorting Technique of hand-sorted naked eyes identification so far.

So far, for above technical problem, people develop the hardware means of hundreds of software identification method, high-speed, high precision, pretreatment is carried out to image, pre-treatment, post-processing technology carry out image border and identify and segmentation, in particularly high-speed automated use, effect is not very desirable, always produces erroneous judgement and generation of failing to judge, does not reach the identification requirement of 100%.

Summary of the invention

The object of this invention is to provide a kind of intelligence feel digital image recognition sorting equipment, solve in prior art, to needing the detection piece differentiating positive and negative, automation can not be realized with low cost in detection, identification, sorting link, in dynamic field, particularly article move in high-speed automated identifying continuously, always produce erroneous judgement and fail to judge, not reaching the problem of 100% identification.

Another object of the present invention is to provide a kind of intelligence feel digital image recognition method for sorting.

The technical solution adopted in the present invention is, a kind of intelligence feel digital image recognition sorting equipment, on delivery track measured piece mobile alignment position above be respectively arranged with at least one group of LED, CCD camera is provided with above the axle center of LED, CCD camera is connected with PC by USB connecting line, and the two ends of USB connecting line are respectively arranged with a demagnetization ring; PC is connected with touch-screen, PC is connected with PLC, the output of PLC is corresponding with multiple magnetic valve respectively to be connected, and each magnetic valve connects with corresponding purge valve and (or) manipulator again, and each purge valve and (or) manipulator are arranged on delivery track side.

Another technical scheme of the present invention is, a kind of intelligence feel digital image recognition method for sorting, and depend on above-mentioned structure, step comprises:

Step 1:CCD camera obtains the surface image of measured piece on delivery track one side upward;

Step 2:CCD camera is transferred to PC is carried out buffer memory by the surface image of USB interface by measured piece one side upward;

Step 3:PC machine utilizes the image processing program prestored to obtain view data by buffer memory;

Step 4:PC machine, by image processing program, show that measured piece upper surface belongs to the judgement of positive and negative;

Step 5: judge that upper surface is that the measured piece in front normally moves forward along delivery track, enter next step welding and assembly process; Judge that upper surface is that the measured piece of reverse side then sends rejecting information to PLC by PC;

The action of step 6:PLC controller operation magnetic valve, controls purge valve and blows, blow reverse side product off, or reverse side product is sorted out delivery track by PLC manipulator by corresponding inflatable mouth.

The invention has the beneficial effects as follows, achieve and the Intelligent Recognition of detection piece positive and negative 100% is detected and sorting, save cost of labor more than 70%, enhance productivity more than 660%, operation and calculating are simply, identify sorting quick and precisely, reliable and stable, be widely used in industrial products and automatically specially pick streamline, be particularly useful for low-voltage circuit breaker to the automatic identification of contact (two gold) and sorting.

Accompanying drawing explanation

Fig. 1 is the structural representation of apparatus of the present invention;

Fig. 2 is the air blowing sorting partial schematic diagram in apparatus of the present invention;

Positive and negative when Fig. 3 apparatus of the present invention measured piece is contact arranges schematic diagram;

Positive and negative when Fig. 4 apparatus of the present invention measured piece is two gold arranges schematic diagram;

Fig. 5 is the schematic diagram that the touch-screen in apparatus of the present invention shows and operation subregion sets.

In figure, 1.CCD camera, 2. demagnetization ring A, 3. demagnetization ring B, 4.USB connecting line, 5.PC machine, 6.LED lamp, 7. delivery track, 8. measured piece, 9. touch-screen, 10. control box, 11. lead-out terminals, 12.PLC controller, 13. singlechip control panels, 14. power supply plug wires, 15. holding wires, 16. magnetic valves, 17. purge valve, 18. manipulators, 19. power supply wirings, 20. Serial Port Lines, 21. left detection zones, 22. main detection zones, 23. secondary detection zones, 24. right detection zones, 25. inflatable mouth one, 26. inflatable mouths two, 27. inflatable mouths three

In addition, 22. positive and negative viewing areas, 221. verso images, 222. direct pictures,

23 surveyed areas, 231. left detection windows, 232. main detection windows, 233. secondary detection windows, 234. right detection windows,

24. setting parameter regions, 241. setting buttons, 242. air blowing window areas setting, 243. surveyed areas setting, 244. put in place region setting, 245. available gray-scales setting,

25. system viewing areas, 251. effective ratio gray scale displays, 252. effective stain displays, 253. system current states and positive and negative display, 254. air blowing states, effectively air blowing number and secondary air blowing digital display are shown,

26. operating areas.

Detailed description of the invention

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.

Apparatus of the present invention detected object (measured piece 8) mainly refers to the contact (two gold) in low-voltage circuit breaker, also can be other products being similar to contact (two gold) shape, referred to as product in following text; When contact (two gold) is fed forward on delivery track 7, the positive and negative of product can only have one to face up, in following text by anti-supine product referred to as reverse side product, by heads product referred to as front product.

With reference to Fig. 1, intelligence of the present invention feels that the structure of digital image recognition sorting equipment is, comprise the delivery track 7 together carrying out transmitting for different types of contact (two gold) being converged sequence, on delivery track 7, conveying has measured piece 8(embodiment to be contact or two gold), above the position of measured piece 8 mobile alignment, be respectively arranged with at least one group of LED 6(preferentially select dome type LED, low angle annular LED light source or the arrangement of lamp battle array, altogether >=1 group, front and back), each LED 6 is provided with independent power supply wiring 19, axle center (or irradiating center) top of LED 6 is provided with CCD camera 1, CCD camera 1 is made a video recording for the contact (two gold) under the range of exposures to LED 6 on delivery track 7, CCD camera 1 is by USB connecting line 4 and PC 5(or industrial computer) be connected, for sending the photographed images of contact (two gold) to PC 5 information process unit, the two ends of USB connecting line 4 are respectively arranged with a demagnetization ring (as the demagnetization ring A2 in Fig. 1 and demagnetization ring B3, the interference of the magnetic field that produces of welding operation to image, data and the control signal in transmitting is eliminated in its effect, avoids the erroneous judgement and the maloperation that cause PC 5 and blowing action), PC 5 is connected with power supply additionally by power supply plug wire 14, PC 5 is connected with touch-screen 9(as man-machine interface), touch-screen 9 is set with the operation interfaces such as recognition detection parameter, detection window, sorting air blowing setting parameter, the operational circumstances of PC 5 in conjunction with touch-screen 9 and the information of photographed images, by the image processing program identification prestored judge contact (two gold) upward one side positive and negative state, PC 5 is connected with PLC 12 by Serial Port Line 20, or PC 5 is connected with control box 10 input by Serial Port Line 20, control box 10 is connected with singlechip control panel 13 by holding wire 15, and control box 10 output is connected with PLC 12 by lead-out terminal 11 again, the output of PLC 12 is corresponding with multiple magnetic valve 16 respectively to be connected, each magnetic valve 16 connects with corresponding purge valve 17 and (or) manipulator 18 again, each purge valve 17 and (or) manipulator 18 are arranged on delivery track 7 side, single-chip microcomputer and PLC 12 provide the instruction of blow off contact or two gold, the effect of purge valve 17 is blown off by reverse side product (or adopting manipulator 18 to be removed in the contact of reverse side or two gold), ensure that front product enters down one processing step by delivery track 7, realize automatically identifying and sorting.

As shown in Figure 2, sorting part of blowing coordinates with recognition detection window arranges, its structure is, throughput direction along delivery track 7 is furnished with four detection zones successively, left detection zone 21 successively, main detection zone 22, secondary detection zone 23 and right detection zone 24, the set-up mode of these four detection zones is arranged in order along orbit transports direction, these four detection zones are all under the image pickup scope of CCD camera 1 and the range of exposures of LED 6, pass through the image corresponding left detection window 231 that be presented at display of contact (two gold) towards upper surface of these four detection zones successively, main detection window 232, in secondary detection window 233 and right detection window 234, outside the track of main detection zone 22, secondary detection zone 23 and right detection zone 24 position, correspondence is provided with inflatable mouth 1, inflatable mouth 2 26 and inflatable mouth 3 27 respectively, then, inflatable mouth 1, inflatable mouth 2 26 and inflatable mouth 3 27 are communicated with respective purge valve 17 respectively, and three groups of purge valve 17 are connected with controlling organization respectively by magnetic valve 16.Sorting part of blowing coordinates with these four detection zones, for being blown off by reverse side product to shaking in dish, so that Posterior circle of resequencing detects and sorting.

Intelligence feels that the detected object of digital image recognition sorting equipment is measured piece 8, and the positive and negative of measured piece 8, through special processing, makes the difference of positive and negative feature more obviously, more be convenient to Intelligent Recognition.

As Fig. 3, positive and negative set-up mode when measured piece 8 is contact is, the front of contact is set to smooth surface, as the left side one in Fig. 2 arranges; And the reverse side of contact is made into grid, pit, convex surface or concave, as the right side four in Fig. 2 arranges.

As Fig. 4, positive and negative set-up mode when measured piece 8 is two gold is set in the front of two gold smooth surface or smooth surface be carved with the word with two golden same color, and the band word 16 as row in left side in Fig. 3 represents the specification of two Nintaus face and two golden 16A; And the reverse side of two gold is printed with the word of black or other obvious distinct color, as two row on right side in Fig. 3.

The recognition principle of apparatus of the present invention is, for the measured piece 8 be applied in different breaker, first obtain measured piece 8 image by CCD camera 1, be transferred to PC 5 by USB connecting line 4 again and carry out buffer memory, the image processing program prestored in PC 5 obtains view data by buffer memory, image processing program draws positive and negative conclusion by the related algorithm preset, if judge that conclusion faces up, then measured piece 8 smooth along delivery track 7 before give; If judge conclusion be reverse side upward, then PC is sent by RS-232 interface and rejects instruction to PLC 12, PLC 12 exports control signal and starts blowing action by magnetic valve 16 and purge valve 17, anti-supine measured piece 8 is blown down by air-flow, or PLC 12 exports control signal and starts manipulator 18 by magnetic valve 16 and removed by reverse side product, leaves delivery track 7.

PC 5 inside in apparatus of the present invention prestores image processing program (detection recognizer).With the supporting touch-screen 9 of PC 5 for realizing showing and the function (or adopt keyboard-mouse to coordinate display screen to operate in addition, embodiment adopts touch-screen 9 to make operation more simple and practical) of touch operation.

With reference to Fig. 5, the action button in touch-screen 9 and image display area arrange and operation principle is:

1) contact (two gold) is transported to below CCD camera 1 along delivery track 7, its image is presented in touch-screen 9, the viewing area of this touch-screen 9 is divided into five regions altogether, the i.e. positive and negative viewing area 22 of product, the surveyed area 23 of product positive and negative, identify and blow sorting setting parameter region 24, the system viewing area 25 of parameter current and state, operating area 26, concrete setting is:

Positive and negative viewing area 22: be included in the product verso images 221 of movement on delivery track 7 or the realtime graphic of direct picture 222;

Surveyed area 23: comprise left detection window 231, main detection window 232, secondary detection window 233 and right detection window 234; Four detection windows are along delivery track 7 direction word order.

Setting parameter region 24: comprise for detection window length setting button and setting coordinate button 241, air blowing window area setting 242, the setting of left surveyed area and the right surveyed area setting 243 to surveyed area 23, put in place region setting 244, surveyed area precision set, left side available gray-scale and right side available gray-scale setting 245.

System viewing area 25: comprise effective ratio gray scale display 251, the effective stain display 252 in available gray-scale ratio, stain ratio, the effective stain in the left side and the right, system current state and positive and negative display 253, air blowing state, effectively air blowing number and secondary air blowing digital display show 254;

Operating area 26: comprise log off, parameters, blowing in the left side, blows in the right, start to detect and console switch and the button such as reset switch.

Intelligence of the present invention feels digital image recognition method for sorting, utilizes above-mentioned intelligence to feel digital image recognition sorting equipment, can to the detection of moving target under static background, comprises pre-setting and actual detection removes two parts,

First, pre-setting before detecting, comprises and carries out feature setting to product positive and negative, the optimum configurations of CCD camera 1, each operating parameter on touch-screen 9 is arranged and the inside of PC 5 prestores recognizer (image border digitized processing);

1, product positive and negative is set, before described, no longer repeat at this.

2, the parameter of CCD camera 1 is set

Make the telecentric lens of CCD camera 1, central region and target surface center superposition, target surface center >=product surface, pixel gray scale is that 20-30, parameter CCD transfer function meets following formula:

$H\left(w\right)=\frac{1}{p}\underset{-\frac{p}{2}}{\overset{\frac{p}{2}}{\text{\∫}}}h{\left(x\right)e}^{\mathrm{jkx}}\mathrm{dx}=\frac{1}{p}\underset{-\frac{a}{2}}{\overset{\frac{a}{2}}{\text{\∫}}}{\mathrm{le}}^{\mathrm{jkx}}\mathrm{dx}=\frac{a}{p}\frac{\mathrm{Sin}\left(\frac{\mathrm{ka}}{2}\right)}{\left(\frac{\mathrm{ka}}{2}\right)},$

P is wherein image space-between, and a is pixel dimension, and k is angular frequency, and h (x) is unit impulse response, e ^jkxit is exactly the j*k*x power of e, j is ' limit ' in signal processing system, namely limit determines the operational mode of signal, be similar to and determine that this signal is from the conduct mode passing to there here, as can select from Shanghai to Shenzhen walking, by bike, by bus, by train, the mode such as to go by air, jkx is exactly the product of limit parameter * angular frequency * impulse response frequency.

The micro-lens visual field (mm)= ^y/ _e, Y is wherein CCD value, and choosing value scope is 3.6-7.4; E is wherein optics speed, and choosing value scope is 2-4.5;

The functional relation of dynamic instantaneous pixel gray scale and electronic shutter time and illuminance is:

G=f (I, t), G are wherein pixel gray scale, and I is illuminance, and t is the electronic shutter time;

Electronic shutter Su Du≤2-2.5 times of product translational speed, in addition, electronic shutter speed equals individual processing speed of image pick-up card, to ensure higher resolution ratio and definition.

3, each operating parameter of touch-screen is set, determines product Kinematic Positioning surveyed area, divide detection and localization window, detect.

Edge, positioning product surrounding, guarantee that present analysis region is at product surface, need delivery track 7 during owing to specifically implementing as surveyed area image background, and the surperficial gray scale of this background and product is more or less the same, therefore need to determine whether product is in surveyed area dynamically, then region surface is analyzed.

Kinematic Positioning surveyed area is set, setting trunnion axis X and vertical axis Y, with the Nogata frame window of target surface Central Symmetry pull-out 1.1-0.98 times of product area, detection window (or Nogata frame window) quantity >=1 arranged, (preceding embodiment is set to four detection windows, and the detection window in fact arranged is less than four also can realize normal detection, but arranging four detection windows adapts to complicated detected status more, reliability is higher, beyond doubt optimal case).

Detect and judge: when image setting surveyed area square frame window four limit all detect to have different gray value with background colour time, such as contact surface can be higher than the surperficial gray scale of delivery track 7 usually, then judge that contact enters surveyed area, the setting and the positive and negative deterministic process that detect opportunity are:

3.1) surveyed area 23 comprises left detection window 231, main detection window 232, secondary detection window 233 and the right detection window 234 of monitoring delivery track 7, when two products are from beginning to end, software intelligence can identify simultaneously at left detection window 231 and secondary detection window 233 and processes, and the marginal explanation realizing continuous product detects.

3.2) for each above-mentioned detection window, when image setting surveyed area >=2/3 square frame detect to have different gray value with background colour time (two gold surface can be higher than the surperficial gray scale of delivery track 7 usually), judge that two gold enters surveyed area.For determining whether when product afterbody enters the gray value of left detection window 231 and background colour has the different time to judge to enter, carrying out product and putting in place to detect and start.

3.3) left detection window 231 pairs of products be detected as on-line checkingi, left detection window 231 and secondary detection window 233 all adopt the effective pixel points that puts in place to detect, the area in the effective pixel points that puts in place region sets according to practical adjustments, basis for estimation is area >=100 pixel of setting regions, namely start product positive and negative at product introduction to detect, the front end, region of main detection window 232 is provided with inflatable mouth 1, do not detect when product head enters left detection window 231, when the afterbody of product enters left detection window 231, effective pixel points area >=100 pixel of afterbody, provide product and run the signal put in place, main detection window 232 starts to detect, main detection window 232 is according to available gray-scale point (or stain, the points such as shade) check product upper surface, be judged as front product normal through, when being judged as reverse side product, the head of product has entered secondary detection window 233, the front end, region of this secondary detection window 233 is provided with inflatable mouth 2 26, when head effective pixel points area >=100 pixel of product, provide product and run the signal put in place, start to detect, be judged as that reverse side is then blown upward, blow reverse side product off, be judged as facing up, do not blow, front product normal through, when entering right detection window 234, the region of right detection window 234 is provided with inflatable mouth 3 27 in addition, and system (finally) judges again, is judged as that reverse side is then blown upward, blows reverse side product off, is judged as facing up, does not blow, front product normal through.

3.4) secondary detection window 233 judges whether product leaves inflatable mouth, criterion is: 50% of the effective pixel points >=detection window vertical height of product, horizontal effective pixel points >=tri-point is connected (or being partitioned into several rows of cloth) in a row, then product have passed through inflatable mouth and leaves, and leaves left detection window mouth region and does not detect (detect and stop).

3.5) set the lower limit of reverse side product valid pixel, the black display of left detection window 231 reverse side product frame, the lower limit of effective stain (shade) (being assumed to be 2169) >=valid pixel, is judged as reverse side product, and this is first reverse side Rule of judgment; The main detection window 232 of product introduction, during the length distance of stain (shade) >=50% main detection window of the product upper bottom display in main detection window 232, be judged as reverse side product, this is second reverse side Rule of judgment; Detect the effective ratio K of reverse side, meet the effective ratio K of reverse side, be judged as reverse side product, this is the 3rd reverse side Rule of judgment.

Any one (the referring to above-mentioned 3.2), 3.3 in above three conditions can be met in practical application), 3.4) any one), be namely judged as reverse side product; Otherwise be judged as front product.

3.6) product is detect in real time through right detection window 234, set the lower limit of this right detection window 234 reverse side product valid pixel, this lower limit is set as 2000, dynamic monitoring Product checking window valid pixel >=lower limit, be judged as reverse side product, otherwise be front product.

In fact, above-mentioned 3.3), 3.4), 3.5), 3.6) be all judging positive and negative, criterion of these four kinds of criterions any one or its combination, and right detection window 234 is exist as the needs of making provision against emergencies substantially, according to above-mentioned structure, apparatus of the present invention just can complete detection at left detection window 231 or main detection window 232, just be in the consideration of reliability, just in case the inflatable mouth above or action of manipulator is not in place causes actual removing to occur error, so need finally carry out one-time detection again and remove for the last time at right detection window 234.

Image edge processing recognizer in PC is arranged; The enhancing process of photographed images comprises the process of removing background, impurity, the fuzzy and very big area of photographed images dynamic gray level, ask for the feature digitized processing such as photographed images concave-convex surface, texture, pit, character, color, height gray scale converges the ratio of the GTG area of the edge image of connection, for judging the positive and negative of product;

4, the pretreatment of product image is obtained by CCD camera 1 shooting

The image of input Kinematic Positioning surveyed area background and contact (two gold), remove the impurity of Kinematic Positioning surveyed area background and contact (two gold) positive and negative:

D (x, y) * h (x, y)=M (x, y)-S (x, y), the * in formula represents convolution algorithm,

The actual original image of D (x, y) wherein for obtaining, M (x, y) is dynamic realtime input degraded image, and S (x, y) is impurity noise figure picture, and h (x, y) represents the spatial description of degenrate function;

The image moved across covers background, and automatic shield is not at the row of calculating;

5, image deblurring and the very big area handling procedure of dynamic gray level, specifically comprises:

5.1) contact (two gold) dynamically deblurring process, with reference to following formula:

Mx＝V*sina×t÷W×D，My＝V*cosa×t÷W×D，

Mx is wherein landscape blur degree, and My is longitudinal fuzziness, Mx≤2 pixel, My≤2 pixel; V is contact (two gold) speed of service, and a is the operation angle of the relative CCD camera 1 of contact (two gold), and t is the time, and W is actual scene width, and D is picture traverse;

5.2) process of the very big area of dynamic gray level

Remove impurity image, comprise the maximum gross area (i.e. the area upper limit) of background in study detection window, front or reverse side collection; Dynamic use configuration scope surrounds the impurity (as flaw, stain, burr, defect etc.) of background, product surface, is registered in the current maximum gross area, is shielded or remove after calculating;

The maximum gross area of the impurity such as flaw, stain, burr, defect of dynamic operation generation does not below exceed the maximum gross area of history, without the need to calculating and registration;

Otherwise registration after needing accumulation to calculate, carry out the renewal of background, again will shield or remove impurity, the area upper limit is A _n, R _nthe set of neighbor average threshold, R _nin pixel number be:

$A_n=\underset{\text{x}}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}f(x,y),(x,y)\∈R_n;$

6, the image border digitized processing recognizer of PC 5 inside is set

Ask for setting edge image effective GTG parameter area ratio, judge that product upper surface is as front or reverse side with this.

Contact (two gold) selects Tuscany image border recognizer and segmentation (also can select the related algorithm except Tuscany edge algorithms); Canny edge detection algorithm is adopted to analyze surveyed area image, obtain the border of product and concavo-convex, texture, pit, character, the color status analysis result of surveyed area interior surface, Canny edge detection algorithm carries out the step that judges successively: read in the positive and negative that image-Gaussian smoothing-gradient calculation-non-maxima suppression-double-threshold comparison obtains image border-border and to track and low gray scale Hybrid connections image border-Hybrid connections image border binaryzation-effectively GTG area-delimitations image gray-scale level area ratio-judge entire image, below detailed description:

6.1) image-Gaussian smoothing image is read in

Carry out filtering with Gauss's first derivative wave filter to reading in image, namely carry out filtering with Gaussian function to image f (x, y) and obtain smoothing image data matrix, (x, y) is image pixel coordinates respectively, then have:

s(x,y)＝f(x,y)*G(x,y,σ)，

Above formula is Gaussian filter function G (x, y, σ) with original image f (x, y) through carrying out process of convolution, obtains a level and smooth image S (x, y) after filtering; σ in G (x, y, σ) is specified value deviation, and image uses and smoothly, reduces noise with this with specified value deviations Gaussian filter,

Gaussian filter function is: $G(x,y,\mathrm{\σ})=\frac{1}{{2\mathrm{\π\σ}}^2}\exp (-\frac{1}{2\mathrm{\σ}}(x^2+y^2)),$

G (x, y, σ) is two-dimensional Gaussian function, a direction n is the first directional derivative of G (x, y), then:

$G_n=\frac{\∂G}{\∂n}=n\▿G,$ Wherein $n=\left[\begin{array}{c}\cos \mathrm{\θ}\\ \sin \mathrm{\θ}\end{array}\right],\▿G=\left[\begin{array}{c}\frac{\∂G}{\∂x}\\ \frac{\∂G}{\∂y}\end{array}\right],$

N in formula is direction vector, and ▽ G is gradient vector, by image f (x, y) and G _nmake convolution, change the direction of n, then G simultaneously _n* n when f (x, y) obtains maximum is exactly the direction being orthogonal to Edge detected; It is level and smooth that image uses the Gaussian filter with specified value deviations, and σ span is 0.3-3, thus can reduce noise;

First image is used the smoothing process of Gaussian function, then by the maximum determination marginal point of first differential; The maximum that the zero cross point of second dervative not only correspond to first derivative also correspond to the minimum of first derivative, and that is, the point (strong edge) that grey scale change is violent and grey scale change slowly point (weak edge) all correspond to second dervative zero cross point; By using two threshold values to detect strong edge and weak edge respectively, and and if only if when strong edge is connected with weak edge, weak edge just can comprise in the output, and therefore, Canny algorithm is not easy the interference by noise, real weak edge can be detected.

6.2) gradient magnitude of image and gradient direction after calculation of filtered

Difference compute gradient amplitude M and deflection A according to the following formula:

M＝||f(x,y)*G(x,y,σ)||，

$A=\frac{f(x,y)*G(x,y,\mathrm{\σ})}{\left|\right|f(x,y)*G(x,y,\mathrm{\σ})\left|\right|},$

Two these symbols erected are called norm, in fact it is by the mapping of linear normed spaces to nonnegative real number, and in linear normed spaces, it is for the point in representation space and the distance between initial point, the distance of point-to-point transmission is also represent by the norm of the difference of 2, norm the condition that meets have:

(1) || x||>=0, and || x||=0 and if only if x=0,

(2) || ax||=|a|*||x|| wherein a is the number in number field corresponding to linear space,

（3）||x+y||<=||x||+||y||；

Conversely, the mapping meeting above condition in linear normed spaces all can be described as norm.

M is wherein gradient magnitude (mould of gradient vector), and M reflects the edge strength of image; A is deflection, and A reflects the direction at edge, obtains the direction that M and the deflection A that obtains local maximum just reflects edge;

6.3) non-maximum is carried out to gradient magnitude and face upward system

Select the vicinity points of gradient magnitude contrast, carrying out non-maximum suppression processing procedure is:

Algorithm follows the trail of the top of the ridge of the ridge band of all refinement magnitude image, and be not set to zero in the pixel at the top of ridge by all, only retain the maximum point of amplitude localized variation, to provide a fine rule in the output, produce the edge that refinement width is 1 pixel width, pixel larger than neighbor point gradient magnitude on gradient direction for gradient magnitude is detected as marginal point, and non-maximum image border point is and makes M obtain the point of local maximum on deflection A, as shown in the formula:

N(i,j)＝NMS(M(i,j),c(i,j))＝Sector(θ[i,j])，

Non-maxima suppression magnitude image is N (i, j), gradient angle is (θ [i, j]), sector value c (i, j), and the label of four sectors is respectively 0 to 3, and four kinds of correspond in 33 neighborhoods may combine.

6.4) double-threshold comparison obtains image border

Obtain the high threshold threshold value of Canny algorithm by maximum entropy algorithm and obtain threshold ones accordingly, the typical method reducing false amount of edge is to N (i, j) threshold value is used, all values lower than threshold value is composed null value, the result of non-maxima suppression amplitude being carried out to thresholding is an image I (i, j) edge array, the edge array obtained after thresholding still has false marginal existence, reason is the existence of threshold value too low (false correct) and shade, contrast on border is weakened, or threshold value T obtains Tai Gao and causes partial contour to lose (false error by mistake), suitable threshold value is selected to be difficult, in order to address this problem, Canny proposes a kind of dual threshold method, namely cumulative statistics histogram is first utilized to obtain a high threshold T ₁, and then get a Low threshold T ₂, setting

If the response of picture signal is greater than high threshold, so it must be edge; If lower than Low threshold, so it is edge scarcely; If between Low threshold and high threshold, just see whether its eight adjacent pixels are greater than the edge of high threshold, if had, so it is edge, otherwise it is not edge;

6.5) border tracks and low gray scale Hybrid connections image border

One, the top fine rule of the ridge that the marginal point that above non-maximum faces upward system is provided tracks as border; Face upward due to filtering and non-maximum and manufacture fine rule fracture, discontinuous, take the effective low gray scale of the 3-11Pix of Edge detected width around a fine rule to be surrounded to connect, form border and track and low gray scale Hybrid connections image border;

Effective low gray scale refers to and is greater than 25Pix and occupies certain proportion in 0-255 GTG, and occupies 10% to 60% scope of effectively most high gray, and preferable range is 20%-45%; The removal that certain proportion does not reach more than 10% ratio will not be occupied in 0-255 GTG lower than 25Pix,

Effective low gray scale=available gray-scale+(10%-60%) adjustment portion gray scale recently;

6.6) Hybrid connections image border binaryzation, ask for the rule judging product positive and negative:

The pixel of image own is made up of the pixel of different gray scale, gray value itself exists, in computational process, utilize minimum and maximum function to deposit the very big gray value H of periphery and minimum gray value L respectively, GTG thresholding is variable partition on software interface, namely setting grey decision-making is more than or equal to this is worth this grey decision-making again set 255, being less than this thresholding grey decision-making, to obtain set be again 0, the binary image that such realization needs image, after completing following edge detection, utilize binaryzation GTG processing mode, provide three kinds of judgment modes and corresponding standard:

The first scheme

The height gray scale area of the Hybrid connections image border after binaryzation GTG or girth; Judge the height gray scale area parameters formula of the Hybrid connections image border of product positive and negative, be applicable to coloured and silvery white face that product positive and negative is set, first obtain the highest gray scale H of photo current and minimum gray scale L:

$H=\frac{1}{K}\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{H}_k,L=\frac{1}{K}\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}{L}_k,$

The number of each extreme value that K representative calculates, the discrete extreme value quantity namely obtained, as we can obtain the highest inconsistent minimum gray scale in zones of different, waveform can be zigzag, in order to remove zigzag, also just similar smothing filtering is averaged as average extreme value to local extremum, and two formula are namely sued for peace the formula be averaged.

A suitable GTG threshold is on average set according to the difference of two gray scales, calculate binarized pixel area SH, SL after surveyed area binaryzation, gradient image after process is connected with dual threshold algorithm, the scope of effectively most high gray M is 10%-60%, M preferable range is 20%-45%, obtains edge image;

For an image-region, its area (SH or SL) is that in image-region, valid pixel is counted, then have:

SL=Sum(L ₁+L ₂+L ₃...+L _K)，

SH=Sum(H ₁+H ₂+H ₃...+H _K)，

A=SL+SH，

Edge surface dust stratification rank:

When judging for dim spot or shade, K=SL/SL+SH;

When judging for bright spot or bright shadow, K=SH/SL+SH;

SH and SL is wherein respectively by valid pixel quantity and represent effective pixel area value; Effective pixel area then represents the pixel quantity of whole detection square frame, because therefore the size of each pixel necessarily represents effective pixel area by number of pixels is the simplest mode,

According to product positive and negative image gray-scale level area ratio detection coefficient K, judge edge image positive and negative, wherein 60% >=K >=10%(is used for dim spot or shade judges) be judged as reverse side, when K is more than or equal to reverse side mark ratio, then judge that now detection faces is as product reverse side, otherwise now detection faces is product front.

First scheme

Judge the height gray scale area parameters formula of the Hybrid connections image border girth of product positive and negative, be applicable to the silvery white face that contact (two gold) positive and negative is set:

Selected in 1Pix to 11Pix scope by the width of Hybrid connections image border (one, the top fine rule of the ridge) greyscale pixel of product positive and negative, statistical definition arranges the height gray scale of the Hybrid connections image border girth of contact (two gold) positive and negative:

$L_H=\underset{K=1}{\overset{Q}{\mathrm{\&Sigma;}}}{\mathrm{SH}}_K,L_L=\underset{K=1}{\overset{Q}{\mathrm{\&Sigma;}}}{\mathrm{SL}}_K,$

Q is wherein borderline pixel number, SH _kfor the highest gray scale on border, SL _kfor the minimal gray on border, perimeter L is count in the border of region R,

Hybrid connections image border girth GTG detection ratio coefficient formula is:

When judging for dim spot or shade, K1=L _l/ (L _h+ L _l),

When judging for bright spot or bright shadow, K1=L _h/ (L _h+ L _l)

K1 preferable range is 20%-45%, L _lrepresent the low gray scale of Hybrid connections image (one, the top fine rule of ridge) as rim circumference, L _hrepresent the high gray scale of Hybrid connections image (one, the top fine rule of ridge) as rim circumference,

According to product positive and negative image border girth GTG detection ratio COEFFICIENT K 1, judge edge image positive and negative, wherein 45% >=K1 >=20%(is used for dim spot or shade judges) be judged as reverse side; When K1 is more than or equal to reverse side mark ratio, then judge that detection faces is as reverse side herein, otherwise now detection faces is front.

The third scheme

Judge two golden reverse side ink-jet character, edge, the Hybrid connections image border height grey parameter formula carved characters in front:

Edge, the Hybrid connections image border gray scale of the two golden reverse side ink-jet character of statistics, literal effective ratio gray scale area (dark face gray scale area) >=50% is had to can be regarded as dark face available gray-scale area, the gray scale area on the right namely in histogram, obtains current area n by cumulative for current available gray-scale

(having literal gray scale+degree of regulation)≤background gray scale, accuracy of detection is 15.50Pix;

Then reverse side ink-jet character judgment formula is: the Kn=n/ gross area >=50%,

Wherein, the gross area=dynamically square frame x value × dynamically square frame y value, the gross area, after debugging is determined, dynamically keeps certain relatively.

Two golden reverse side ink-jet character is the criterion of reverse side: Kn >=50%;

Statistics two Nintaus carve characters in face edge, the Hybrid connections image border gray scale of symbol, effective front >=800, face of carving characters, and reverse side is 0, and the pixel of the setting value that rolls off the production line is 100, accuracy of detection 20-40Pix,

Two Nintaus symbol of carving characters in face for the criterion of reverse side is: the effective front >=866Pix in face of carving characters; Reverse side is 0Pix, reverse side ink-jet character, judges by bright gray scale.

In sum, intelligence feel digital image recognition method for sorting of the present invention, utilize above-mentioned intelligence feel digital image recognition sorting equipment, step comprises:

Pre-set step, be included in setting, PC 5 inner pre-stored image edge digitlization processing and identification program that product positive and negative carries out the relevant Intelligent Recognition sorting parameter on feature setting, the optimum configurations of CCD camera 1, touch-screen 9;

Actual detecting step

Step 1: to be shaken dish or conveyer belt by delivery track 7() product is transported to the irradiation position place of CCD camera 1, CCD camera obtains the surface image of measured piece 8 on delivery track 7 one side upward;

Step 2:CCD camera 1 is transferred to PC 5 is carried out buffer memory by the surface image of USB interface by product one side upward;

Step 3:PC machine 5 utilizes the image processing program prestored to obtain view data by buffer memory;

Step 4:PC machine 5 is by the CANY Algorithm Analysis in image processing program, and combining image gray scale area ratio, show that product upper surface belongs to the judgement of positive and negative;

Step 5: judge that upper surface is that the product in front normally moves forward along delivery track 7, enter next step welding and assembly process; Judge that upper surface is that the product of reverse side then sends rejecting information to PLC 12 by PC 5 by RS232;

Step 6:PLC controller 12 operates magnetic valve 16 action, control purge valve 17 to blow, reverse side product (or reverse side product is sorted out delivery track 7 by PLC 12 manipulator 18) is blown off by corresponding inflatable mouth, the product that this is blown off (or sorting out) drops to and shakes in dish, and rearrangement identifies sorting.

Embodiment 1

The inventive method is used for the two gold of detection of dynamic and has literal identification, test item and parameter respectively: in original graph, utilize Pixel Dimensions to be 752 × 480Pix; Frame per second is the CCD of MAX60F/S, have literal by detection of dynamic and program to two gold of motion continuously, border must be arrived track and low gray scale Hybrid connections image border, to track on border and low gray scale connects in combination chart, the standard variance of Gaussian filter is set to 3, detects edge Breadth Maximum and is set to 11Pix; In binary picture, edge width selects >=3, and in histogram, utilizing detection method of the present invention to obtain two gold has literal detected parameters: effectively pixel number is 56672Pix, and the highest available gray-scale is 255Pix, and minimum available gray-scale is 94Pix; High gray scale area SH is 42852Pix, and low gray scale area SL is 13820Pix, detection ratio K is 0.2-0.42, utilizes detection ratio K through the detection of 1,500 ten thousand products, and two gold has 1,000,000 literal qualification rates to be 100%.

Embodiment 2

The inventive method is used for the two gold of detection of dynamic without literal recognition detection, test item and parameter respectively: in original graph, utilize Pixel Dimensions to be 752 × 480Pix; Frame per second is the CCD of MAX60F/S, to two gold of continuously motion without literal by detection of dynamic and program, border must be arrived track and low gray scale Hybrid connections image border, to track on border and low gray scale connects in combination chart, the standard variance of Gaussian filter is set to 3, and detecting edge Breadth Maximum is 7Pix; In binary picture, edge width selects >=3, in histogram, utilizes detection method of the present invention to obtain two gold without literal detected parameters: effectively pixel number is 56672Pix, and the highest available gray-scale is 248Pix, and minimum available gray-scale is 140Pix; High gray scale area SH is 56464Pix, and low gray scale area SL is 208Pix, detection ratio K is 0.68-0.8, and utilize detection ratio K through the detection of 1,600 ten thousand products, two gold is 100% without 1,000,000 literal qualification rates.

Embodiment 3

The inventive method is used for detection of dynamic large white square silver point reverse side recognition detection, test item and parameter respectively: in original graph, utilize Pixel Dimensions to be 752 × 480Pix; Frame per second is the CCD of MAX60F/S, to the large white square silver point reverse side moved continuously by detection of dynamic and program, border must be arrived track and low gray scale Hybrid connections image border, to track on border and low gray scale connects in combination chart, the standard variance of Gaussian filter is set to 3, and detecting edge Breadth Maximum is 8Pix; In binary picture, edge width selects >=3, in histogram, utilizes detection method of the present invention to obtain the detected parameters of large white square silver point reverse side: effectively pixel number is 48825Pix, and the highest available gray-scale is 240Pix, and minimum available gray-scale is 40Pix; High gray scale area SH is 37971Pix, low gray scale area SL is 10854Pix, detection ratio K is 0.11-0.36, utilizes detection ratio K through the detection of 39,000 ten thousand products, 1,000,000 qualification rates of large white square silver point reverse side are that 100%, 1,000,000 loss and false drop rate are 0.

Embodiment 4

The inventive method is used for detection of dynamic large white square silver point front recognition detection, test item and parameter respectively: in original graph, utilize Pixel Dimensions to be 752 × 480Pix; Frame per second is the CCD of MAX60F/S, to the large white square silver point front of moving continuously by detection of dynamic and program, border must be arrived track and low gray scale Hybrid connections image border, to track on border and low gray scale connects in combination chart, the standard variance of Gaussian filter is set to 3, and detecting edge Breadth Maximum is 5Pix; In binary picture, edge width selects >=5, in histogram, utilizes detection method of the present invention to obtain the detected parameters in large white square silver point front: effectively pixel number is 48825Pix, and the highest available gray-scale is 47Pix, and minimum available gray-scale is 13Pix; High gray scale area SH is 48509Pix, low gray scale area SL is 316Pix, detection ratio K is 0.64-0.89, utilizes detection ratio K through the detection of 49,800 ten thousand products, 1,000,000 qualification rates in large white square silver point front are that 100%, 1,000,000 loss and false drop rate are 0.

Embodiment 5

The inventive method is used for the large white of detection of dynamic long block silver point reverse side recognition detection, test item and parameter respectively: in original graph, utilize Pixel Dimensions to be 752 × 480Pix; Frame per second is the CCD of MAX60F/S, to the large white long block silver point reverse side moved continuously by detection of dynamic and program, border must be arrived track and low gray scale Hybrid connections image border, to track on border and low gray scale connects in combination chart, the standard variance of Gaussian filter is set to 3, and detecting edge Breadth Maximum is 9Pix; In binary picture, edge width selects >=3, in histogram, utilizes detection method of the present invention to obtain the detected parameters of large white long block silver point reverse side: effectively pixel number is 34632Pix, and the highest available gray-scale is 255Pix, and minimum available gray-scale is 63Pix; High gray scale area SH is 24950Pix, low gray scale area SL is 9682Pix, detection ratio K is 0.16-0.36, utilizes detection ratio K through the detection of 59,800 ten thousand products, 1,000,000 qualification rates of large white long block silver point reverse side are that 100%, 1,000,000 loss and false drop rate are 0.

Embodiment 6

The inventive method is used for the large white of detection of dynamic long block silver point front recognition detection, test item and parameter respectively: in original graph, utilize Pixel Dimensions to be 752 × 480Pix; Frame per second is the CCD of MAX60F/S, to the large white long block silver point front of moving continuously by detection of dynamic and program, border must be arrived track and low gray scale Hybrid connections image border, to track on border and low gray scale connects in combination chart, the standard variance of Gaussian filter is set to 3, and detecting edge Breadth Maximum is 4Pix; In binary picture, edge width selects >=3, in histogram, utilizes detection method of the present invention to obtain the detected parameters in large white long block silver point front: effectively pixel number is 34632Pix, and the highest available gray-scale is 91Pix, and minimum available gray-scale is 13Pix; High gray scale area SH is 34460Pix, low gray scale area SL is 172Pix, detection ratio K is 0.64-0.84, utilizes detection ratio K through the detection of 49,800 ten thousand products, 1,000,000 qualification rates in large white long block silver point front are that 100%, 1,000,000 loss and false drop rate are 0.

Embodiment 7

The inventive method is used for detection of dynamic coloured square silver point reverse side Intelligent Recognition detect, test item and parameter respectively: in original graph, utilize Pixel Dimensions to be 752 × 480Pix; Frame per second is the CCD of MAX60F/S, to the coloured square silver point reverse side moved continuously by detection of dynamic and program, border must be arrived track and low gray scale Hybrid connections image border, to track on border and low gray scale connects in combination chart, the standard variance of Gaussian filter is set to 3, and detecting edge Breadth Maximum is 12Pix; In binary picture, edge width selects >=3, in histogram, utilizes detection method of the present invention to obtain the detected parameters of coloured square silver point reverse side: effectively pixel number is 21450Pix, and the highest available gray-scale is 204Pix, and minimum available gray-scale is 45Pix; High gray scale area SH is 11866Pix, low gray scale area SL is 9584Pix, detection ratio K is 0.14-0.32, utilizes detection ratio K through the detection of 69,800 ten thousand products, 1,000,000 qualification rates of coloured square silver point reverse side are that 100%, 1,000,000 loss and false drop rate are 0.

Embodiment 8

The inventive method is used for detection of dynamic white nahlock silver point positive and negative Intelligent Recognition to detect, correction data sees the following form 1, shown in the form of table 2.

Table 1, the inventive method is used for detection of dynamic white nahlock silver point front Intelligent Recognition and detects data

Test item	Detected parameters	Test item	Detected parameters
				Effective pixel element	16764Pix	Effective pixel element	16764Pix
The highest available gray-scale	246Pix	The standard variance of Gaussian filter	16764Pix
				Minimum available gray-scale	58Pix	Detect edge Breadth Maximum	8.0Pix
Edge width is selected	≥3.0Pix	CCD pixel element	752x480Pix
				High gray scale area SH	15928Pix	Detection ratio K	0.85.0.9Pix
Low gray scale area SH	836Pix	1000000 qualification rates	100%
				CCD frame per second	Max60F/s	1000000 loss	0
		1000000 false drop rates	0

Table 2, the inventive method is used for detection of dynamic white nahlock silver point reverse side Intelligent Recognition and detects data

Test item	Detected parameters	Test item	Detected parameters
				Effective pixel element	16764Pix	Effective pixel element	16764Pix
The highest available gray-scale	255Pix	The standard variance of Gaussian filter	3.0Pix
				Minimum available gray-scale	107Pix	Detect edge Breadth Maximum	13Pix
Edge width is selected	≥3.0Pix	CCD pixel element	752x480Pix
				High gray scale area SH	8950Pix	Detection ratio K	0.27-0.58Pix
Low gray scale area SH	7814Pix	1000000 qualification rates	100%
				CCD frame per second	Max60F/s	1000000 loss	0
		1000000 false drop rates	0

Embodiment 9

The inventive method is used for detection of dynamic coloured square silver point front Intelligent Recognition to detect, test item and parameter respectively: in original graph, utilize Pixel Dimensions to be 752 × 480Pix; Frame per second is the CCD of MAX60F/S, to coloured square silver point front of moving continuously by detection of dynamic and program, border must be arrived track and low gray scale Hybrid connections image border, to track on border and low gray scale connects in combination chart, the standard variance of Gaussian filter is set to 3, and detecting edge Breadth Maximum is 5Pix; In binary picture, edge width selects >=3, in histogram, utilizes detection method of the present invention to obtain the detected parameters in coloured square silver point front: effectively pixel number is 21450Pix, and the highest available gray-scale is 86Pix, and minimum available gray-scale is 30Pix; High gray scale area SH is 21168Pix, low gray scale area SL is 282Pix, detection ratio K is 0.68-0.86, utilizes detection ratio K through the detection of 79,800 ten thousand products, 1,000,000 qualification rates in coloured square silver point front are that 100%, 1,000,000 loss and false drop rate are 0.

Another device and method of the present invention, by to NM1 dynamic/static contact, DZ47-60 static contact, volume production checking on probation on the automatic identification of the products such as the two gold of DZ15 dynamic/static contact, NB1 and DZ47 and sorting positive and negative, automatic welding device, effect ideal very, reaches the identification of 100%.

Claims (3)

1. an intelligence feel digital image recognition method for sorting, depend on a kind of intelligence feel digital image recognition sorting equipment, this apparatus structure is,

At least one group of LED (6) is respectively arranged with above the position of upper measured piece (8) mobile alignment of delivery track (7), CCD camera (1) is provided with above the axle center of LED (6), CCD camera (1) is connected with PC (5) by USB connecting line (4), and the two ends of USB connecting line (4) are respectively arranged with a demagnetization ring; PC (5) is connected with touch-screen (9), PC (5) is connected with PLC (12), the output of PLC (12) is corresponding with multiple magnetic valve (16) respectively to be connected, each magnetic valve (16) connects with corresponding purge valve (17) and manipulator (18) again, and each purge valve (17) and manipulator (18) are arranged on delivery track (7) side;

Described measured piece (8) for positive and negative set-up mode during contact be that front is set to smooth surface; Reverse side is then made into grid, pit, convex surface or concave;

Positive and negative set-up mode when described measured piece (8) is two gold is that front is set to smooth surface; Reverse side is then printed with word;

Throughput direction along delivery track (7) is furnished with four detection zones successively, be left detection zone (21), main detection zone (22), secondary detection zone (23) and right detection zone (24) successively, these four detection zones are all under the image pickup scope of CCD camera (1) and the range of exposures of LED (6); Outside the track of main detection zone (22), secondary detection zone (23) and right detection zone (24) position, correspondence is provided with inflatable mouth one (25), inflatable mouth two (26) and inflatable mouth three (27) respectively, inflatable mouth one (25), inflatable mouth two (26) and inflatable mouth three (27) are communicated with respective purge valve (17) respectively, and three groups of purge valve (17) are connected with PLC (12) respectively by magnetic valve (16);

Described PC (5) is connected with control box (10) input by Serial Port Line (20), control box (10) is connected with singlechip control panel (13) by holding wire (15), control box (10) output is connected with PLC (12) by lead-out terminal (11) again

Utilize above-mentioned device, it is characterized in that, the step of the method comprises:

Pre-set, be namely included in setting, PC (5) inner pre-stored image edge digitlization processing and identification program that measured piece (8) positive and negative carries out the relevant Intelligent Recognition sorting parameter on feature setting, the optimum configurations of CCD camera (1), touch-screen (9);

Step 1:CCD camera obtains the surface image of measured piece (8) on delivery track (7) one side upward;

Step 2:CCD camera (1) is transferred to PC (5) is carried out buffer memory by the surface image of USB interface by measured piece (8) one side upward;

Step 3:PC machine (5) utilizes the image processing program prestored to obtain view data by buffer memory;

Step 4:PC machine (5), by image processing program, show that measured piece (8) upper surface belongs to the judgement of positive and negative;

Step 5: judge that upper surface is that the measured piece (8) in front normally moves forward along delivery track (7), judges that upper surface is that the measured piece (8) of reverse side then sends rejecting information to PLC (12) by PC (5);

Step 6:PLC controller (12) operation magnetic valve (16) action, control purge valve (17) to blow, blow reverse side product off by corresponding inflatable mouth, or reverse side product is sorted out delivery track (7) by PLC (12) manipulator (18).

2. intelligence according to claim 1 is felt and be it is characterized in that digital image recognition method for sorting, and described in PC (5) inner pre-stored image edge digitlization processing and identification program, comprise image border digitized processing recognizer, its process is:

6.1) image-Gaussian smoothing image is read in:

Carry out filtering with Gauss's first derivative wave filter to reading in image, namely carry out filtering with Gaussian function to image f (x, y) and obtain smoothing image data matrix, (x, y) is image pixel coordinates respectively, then have:

s(x,y)＝f(x,y)*G(x,y,σ)，

Above formula is Gaussian filter function G (x, y, σ) with original image f (x, y) through carrying out process of convolution, obtains a level and smooth image S (x, y) after filtering; σ in G (x, y, σ) is specified value deviation,

Gaussian filter function is:

G (x, y, σ) is two-dimensional Gaussian function, a direction n is the first directional derivative of G (x, y), then:

wherein

N in formula is direction vector, gradient vector, by image f (x, y) and G _nmake convolution, change the direction of n, then G simultaneously _n* n when f (x, y) obtains maximum is exactly the direction being orthogonal to Edge detected; σ span is 0.3-3;

6.2) gradient magnitude of image and gradient direction after calculation of filtered:

Difference compute gradient amplitude M and deflection A according to the following formula:

M＝||f(x,y)*G(x,y,σ)||，

The condition that meets have:

(1) || x||>=0, and || x||=0 and if only if x=0,

(2) || ax||=|a|*||x|| wherein a is the number in number field corresponding to linear space,

(3)||x+y||<＝||x||+||y||；

M is wherein gradient magnitude, and A is deflection;

6.3) non-maximum is carried out to gradient magnitude and faces upward system:

Select the vicinity points of gradient magnitude contrast, carrying out non-maximum suppression processing procedure is:

Algorithm follows the trail of the top of the ridge of the ridge band of all refinement magnitude image, and be not set to zero in the pixel at the top of ridge by all, only retain the maximum point of amplitude localized variation, to provide a fine rule in the output, produce the edge that refinement width is 1 pixel width, pixel larger than neighbor point gradient magnitude on gradient direction for gradient magnitude is detected as marginal point, and non-maximum image border point is and makes M obtain the point of local maximum on deflection A, as shown in the formula:

N(i,j)＝NMS(M(i,j),c(i,j))＝Sector(θ[i,j])，

Non-maxima suppression magnitude image is N (i, j), gradient angle is (θ [i, j]), sector value is c (i, j), and the label of four sectors is respectively 0 to 3, and four kinds of correspond in 3 × 3 neighborhoods may combine;

Wherein, i, j represent row and the row of place array respectively, and the implication of NMS is non-maxima suppression, and Sector represents the sector of circumference herein;

6.4) double-threshold comparison obtains image border:

Obtain the high threshold threshold value of Canny algorithm by maximum entropy algorithm and obtain threshold ones accordingly, the typical method reducing false amount of edge is to N (i, j) threshold value is used, all values lower than threshold value is composed null value, the result of non-maxima suppression amplitude being carried out to thresholding is an image I (i, j) edge array, the edge array obtained after thresholding still has false marginal existence, utilizes cumulative statistics histogram to obtain a high threshold T ₁, and then get a Low threshold T ₂, setting

If the response of picture signal is greater than high threshold, so it must be edge; If lower than Low threshold, so it is edge scarcely; If between Low threshold and high threshold, just see whether its eight adjacent pixels are greater than the edge of high threshold, if had, so it is edge, otherwise it is not edge;

6.5) border tracks and low gray scale Hybrid connections image border:

One, the top fine rule of the ridge that the marginal point that above non-maximum faces upward system is provided tracks as border; Take the effective low gray scale of the 3-11Pix of Edge detected width around a fine rule to be surrounded to connect, form border and track and low gray scale Hybrid connections image border;

Effective low gray scale refers to and is greater than 25Pix, and is in 10% to 60% scope of effectively most high gray,

Effective low gray scale=effective minimum gray scale+(10%-60%) adjustment portion gray scale;

6.6) Hybrid connections image border binaryzation, ask for the rule judging product positive and negative:

The pixel of image own is made up of the pixel of different gray scale, gray value itself exists, minimum and maximum function is utilized to deposit the very big gray value H of periphery and minimum gray value L respectively, GTG thresholding is variable partition on software interface, namely setting grey decision-making is more than or equal to this is worth this grey decision-making again set 255, being less than this thresholding grey decision-making, to obtain set be again 0, realize the binary image that image is needed, after completing following edge detection, utilize binaryzation GTG processing mode, there are two kinds of judgment modes and corresponding standard, respectively:

The first scheme:

The height gray scale area of the Hybrid connections image border after binaryzation GTG or girth; First obtain the highest gray scale H of photo current and minimum gray scale L,

The number of each extreme value that K representative calculates, these two formula are namely sued for peace the formula be averaged,

A suitable GTG threshold is on average set according to the difference of two gray scales, calculate binarized pixel area SH, SL after surveyed area binaryzation, connect with dual threshold algorithm the gradient image after process, the scope of effectively most high gray M is 10%-60%, obtains edge image;

For an image-region, its area SH or SL is that in image-region, valid pixel is counted, then have the height gray scale area parameters formula of the Hybrid connections image border judging product positive and negative:

SL＝Sum(L ₁+L ₂+L ₃...+L _K)，

SH＝Sum(H ₁+H ₂+H ₃...+H _K)，

A=SL+SH, A represent edge area GTG summation,

Edge surface dust stratification rank:

When judging for dim spot or shade, Km=SL/ (SL+SH);

When judging for bright spot or bright shadow, Km=SH/ (SL+SH);

SH and SL is wherein respectively by valid pixel quantity and represent effective pixel area value; Effective pixel area then represents the pixel quantity of whole detection square frame,

According to product positive and negative image gray-scale level area ratio detection coefficient Km, judge edge image positive and negative, wherein 60% >=Km >=10%, judge for dim spot or shade, be judged as reverse side, when Km is more than or equal to reverse side mark ratio maximum, then judge that now detection faces is as product reverse side, otherwise now detection faces is product front;

First scheme:

Selected in 1Pix to 11Pix scope by the width of the Hybrid connections image border greyscale pixel of product positive and negative, statistical definition arranges the height gray scale of the Hybrid connections image border girth of product positive and negative:

Q is wherein borderline pixel number, SH _kfor the highest gray scale on border, SL _kfor the minimal gray on border, perimeter L is count in the border of region R, and R refers to the mean radius of image border girth,

Hybrid connections image border girth GTG detection ratio coefficient formula is:

When judging for dim spot or shade, K1=L _l/ (L _h+ L _l),

When judging for bright spot or bright shadow, K1=L _h/ (L _h+ L _l),

L _lrepresent the low gray scale of Hybrid connections image as rim circumference, L _hrepresent the high gray scale of Hybrid connections image as rim circumference,

According to product positive and negative image border girth GTG detection ratio COEFFICIENT K 1, judge edge image positive and negative, wherein 45% >=K1 >=20%, judge for dim spot or shade, be judged as reverse side; When K1 is more than or equal to reverse side mark ratio maximum, then judge that detection faces is as reverse side herein, otherwise now detection faces is front.

3. intelligence according to claim 1 is felt and be it is characterized in that digital image recognition method for sorting, described in PC (5) inner pre-stored image edge digitlization processing and identification program, and the setting and the positive and negative that comprise detection opportunity judge, its process is:

3.1) surveyed area comprises left detection window (231), main detection window (232), secondary detection window (233) and the right detection window (234) of monitoring delivery track (7), when two products are from beginning to end, software intelligence can identify simultaneously at left detection window (231) and secondary detection window (233) and processes, and the marginal explanation realizing continuous product detects;

3.2) for each above-mentioned detection window, when image setting surveyed area >=2/3 square frame detect to have different gray value with background colour time, judge that two gold enters surveyed area;

3.3) left detection window (231) is detected as on-line checkingi to product, left detection window (231) and secondary detection window (233) all adopt the effective pixel points that puts in place to detect, the area in the effective pixel points that puts in place region sets according to practical adjustments, basis for estimation is area >=100 pixel of setting regions, namely start product positive and negative at product introduction to detect, the front end, region of main detection window (232) is provided with inflatable mouth one (25), do not detect when product head enters left detection window (231), when the afterbody of product enters left detection window (231), effective pixel points area >=100 pixel of afterbody, provide product and run the signal put in place, main detection window (232) starts to detect, main detection window (232) is according to available gray-scale point inspection product upper surface, be judged as front product normal through, when being judged as reverse side product, the head of product has entered secondary detection window (233), the front end, region of this secondary detection window (233) is provided with inflatable mouth two (26), when head effective pixel points area >=100 pixel of product, provide product and run the signal put in place, start to detect, be judged as that reverse side is then blown upward, blow reverse side product off, be judged as facing up, do not blow, front product normal through,

When entering right detection window (234), the region of right detection window (234) is provided with inflatable mouth three (27) in addition, system judges again, be judged as that reverse side is then blown upward, blow reverse side product off, be judged as facing up, do not blow, front product normal through;

3.4) secondary detection window (233) judges whether product leaves inflatable mouth, criterion is: 50% of the effective pixel points >=detection window vertical height of product, horizontal effective pixel points >=tri-point is connected in a row, then product have passed through inflatable mouth and leaves, and leaves left detection window mouth region and does not detect;

3.5) set the lower limit of reverse side product valid pixel, the black display of left detection window (231) reverse side product frame, the lower limit of effective stain >=valid pixel, is judged as reverse side product, and this is first reverse side Rule of judgment; The main detection window of product introduction (232), during the length distance of the main detection window of stain >=50% of the product upper bottom display in main detection window (232), be judged as reverse side product, this is second reverse side Rule of judgment.