CN103034838A - Special vehicle instrument type identification and calibration method based on image characteristics - Google Patents

Abstract

The invention provides a special vehicle instrument type identification and calibration method based on image characteristics, and belongs to the field of image processing and mode identification. The method includes the steps of firstly collecting several special vehicle instrument images according to each type of several special vehicle instruments, and reserving representative images as training samples through manual judgment, secondly carrying out normalization to quality of the images through an image preprocessing method according to each training sample, then extracting a disk of the instrument images and carrying out the size normalization to the instrument images according to radius of the disk, thirdly respectively extracting color features and Gabor texture features of the special vehicle instrument images after the normalization, and finally aiming at all training samples, and building models for each type of instruments through respectively using of the color features and the Gabor texture features as vector quantities. After the feature models of the training samples of each type of instruments are built, with regards to the collected real-time images, the method can also be used to carry out image preprocessing and image quality normalization, respectively carries out mode match aiming at the feature templates of each type of instrument training samples with the two features, and obtains instrument classifying results through a nearest neighbour rule.

Description

A kind of identification of special vehicle meter type and scaling method based on characteristics of image

Technical field

The invention belongs to image processing, area of pattern recognition, mainly for be the image of various dissimilar special vehicle instrument, type to the special vehicle instrument is identified automatically, and in conjunction with the prior demarcation of all types of instrument, provides the position of pointer point of fixity and meter dial.

Background technology

Automatic Measurement Technique has a wide range of applications in the production of special vehicle instrument, such as the special vehicle instrument is need to automatically detect the time.For a long time, the calibration of special vehicle instrument is generally adopted manual read's access method with measurement.Although manual detection is accurate, and very large inconvenient part is arranged.The employee who carries out instrument calibration need to manually check total indicator reading, compares with the standard source numerical data, the error of calculation, saving result, in the pointer quick rotation, can have artificial operate miss and people's the collimation error, and calibrating efficient can be very low.Therefore the calibrating of automatic instrument is in demand.The non-contact automatic detection technology of pointer instrument can realize by image processing techniques, namely utilize imageing sensor to simulate human eye, pointer instrument is carried out image acquisition, then with in the video data transmitting that the gathers intelligence system in the computing machine.The recycling computing machine is analyzed and is identified the Instrument image that collects, and obtains at last gauge pointer result pointed.

In the past in the research for pointer instrument, mainly be the automatic identification for instrument, and mainly be to study and compare for industrial instrument, such as document 1 (Chang Faliang, perhaps talented, Qiao Yizheng, the automatic identification and analysis method of unmanned indicator real-time vision, electronic surveying and instrument journal, 2006.4,20 (2): 35-38.), document 2(Sun Lin, Wang Yongdong, pointer instrument automatic Verification image recognition technology, modern electronic technology [J], 2011.Vol.34 (8): 101-104.) etc.For the research of special vehicle instrument identification lacking very, to the automatic identification of special vehicle meter type still less.The special vehicle instrument is very different than industrial instrument, at first special vehicle gauge pointer point of fixity often is not the central point of instrument disk, secondly the scale of very large special vehicle instrument is often inhomogeneous, has brought very large difficulty also just for the automatic identification of pointer point of fixity and meter dial.If can be for the modeling of particular meter type, so just can automatically identify the meter type of the Instrument image of input, according to the prior demarcation of the type, just can obtain the position of pointer point of fixity, the automatic necessary information of identifying of the instrument such as distribution of scale, thereby be conducive to the automatic reading of special vehicle instrument.

Summary of the invention

Purpose of the present invention, be for be the image of various dissimilar special vehicle instrument, the type of special vehicle instrument is identified automatically, and in conjunction with the prior demarcation of all types of instrument, is provided the position of pointer point of fixity and meter dial.

Technical scheme of the present invention is: based on characteristics of image the special vehicle meter type is identified and modeling.The algorithm that extracts feature mainly combines two kinds of characteristics of image, and one of them is color characteristic, and another one is the Gabor textural characteristics, with these two kinds of characteristic synthetics, then with the method for template matches the special vehicle meter type is classified and identifies.

At first to every type special vehicle instrument, gather some special vehicle Instrument images, pass through artificial judgment, representational image is stayed be used as training sample.For each training sample, by image pre-processing methods such as figure image intensifyings picture quality is carried out normalization; Then extract the disk of Instrument image, according to disc radius Instrument image is carried out size normalization; Respectively the special vehicle Instrument image is extracted color characteristic and Gabor textural characteristics after the normalization.Last for all training samples, set up model take color characteristic and Gabor textural characteristics as vector as every type respectively.Its idiographic flow as shown in Figure 2.

After the characteristic model of the training sample database of every type instrument establishes, for the realtime graphic that collects, picture quality normalization is carried out in the pre-service of same process image, extract the instrument disk and carry out the instrument size normalization, extract respectively color characteristic and Gabor textural characteristics after the normalization, feature templates to the instrument training sample of these two kinds of features and each type carries out respectively pattern match, obtains the instrument classification results for the special vehicle instrument of little type with the arest neighbors rule.Its idiographic flow as shown in Figure 1.

Main research of the present invention has: (1) Instrument image normalization: comprise picture quality normalization, Instrument image size normalization; (2) feature extraction: comprise that color characteristic extracts, the Gabor texture feature extraction; (3) feature modeling and pattern match; (4) according to the demarcation of specific meter type, obtain gauge pointer point of fixity, scale information by meter type.

1, Instrument image normalization

Its fundamental purpose of the normalization of special vehicle Instrument image is to reduce the variation between the different samples in the same meter type, namely strengthens the degree of polymerization in the class.Variation comprises following several in the class of typical special vehicle instrument: the illumination of instrument dial plate, the picture quality of instrument, instrument dial plate size, etc.For making a variation in these classes, at first to carry out normalization to the Instrument image quality, this respect mainly relies on pre-service to reduce noise circumstance difference and the photoenvironment difference of various images as far as possible.Then will carry out size normalization to Instrument image, this respect mainly depends on the extraction of instrument disk and radius.Therefore the normalization of special vehicle Instrument image has comprised several steps, is successively sequentially:

● Instrument image strengthens and except denoising,

● the instrument disk extracts,

● according to the linear normalization to the instrument size of the size of disk

The special vehicle Instrument image is carried out pre-service, briefly is exactly for so that the picture quality difference minimum of image and the template image of test.Pre-service can solve image because fuzzy, the crooked or damaged abnormal conditions of Instrument image that the reasons such as light or shooting angle cause.Here taked to go random noise, stretch to strengthen the contrast of Instrument image by color by medium filtering.Best in effect aspect removal noise and the preservation details two with adaptive median filter again in medium filtering, we adopt the modified adaptive median filter algorithm that combines mean filter and adaptive median filter advantage.And the adaptive neighborhood method of average removal of images that uses band to revise disturbs and noise.

For the coloured image with instrument draws high.We obtain first gray level image corresponding to coloured image, and the minimum gray value g1 in the statistics Instrument image, and gray scale maximal value g2 draws high it gray space of (0,255), so that gray level image is more clear.Here adopt a kind of improved gray scale to draw high method, considers minimum gray value g1, and the corresponding grey scale pixel value number of gray scale maximal value g2 is little to the general image improved action seldom the time.We require g1 and the total number of gray-value pixel corresponding to g2 to be configured to and need to be worth greater than certain.If g1 be total pixel greater than the minimal gray of n (such as n=10), g2 is that total pixel is greater than the maximum gray scale of n (such as n=10).For any pixel, its gray-scale value is gray (x, y), and the gray-scale value that draws high rear correspondence is GRAY (x, y), then has:

$\mathrm{GRAY}(x,y)=\frac{255\×[\mathrm{gray}(x,y)-g_1]}{(g_2-g_1)}---\left(1\right)$

For the coloured image of RGB pattern, the redness of its correspondence, green, blue color draw high and are respectively:

$R(x,y)=\frac{255\×[r(x,y)-g_1]}{(g_2-g_1)}---\left(2\right)$

$G(x,y)=\frac{255\×[g(x,y)-g_1]}{(g_2-g_1)}---\left(3\right)$

$B(x,y)=\frac{255\×[b(x,y)-g_1]}{(g_2-g_1)}---\left(4\right)$

We utilize the method for Hough change detection circular arc to obtain the edge of special vehicle instrument disk.After obtaining the size of instrument disk, we carry out linear normalization to the size of instrument.So-called big or small normalization be exactly test pattern and template image all scaling arrive identical size.The normalization fundamental purpose is the variation for radius size between the image that reduces Real-time Collection and the template instrument types of image.What we did is to crop the outer zone of instrument disk, then the pixel in all instrument disks is normalized into the image that radius is 64 pixel sizes.During linear normalization, only do the conversion of equal proportion in X and Y direction.After the normalization, Instrument image becomes the rectangle of 128x128 size, and the edge of rectangular image is the instrument disk and then, and the pixel in the zone that the instrument disk is outer all is arranged to white pixel.

2, feature extraction

Color characteristic is one of key character of special vehicle instrument.Be different from industrial instrument, the part of some special vehicle instrument is colored, such as Fig. 3, and Fig. 6, the instrument among Fig. 9.Most of instrument is black matrix, such as Fig. 3, and Fig. 5, Fig. 6, Fig. 7, Fig. 8, Fig. 9, Figure 10, Figure 11, the instrument of Figure 12, and the color of dissimilar special vehicle instrument is different.According to our observation, in the special vehicle instrument, red, green, black is most important several color, so we carry out feature extraction for these several colors.Our Instrument image that normalizing is good is divided into the grid of 8X8, and each grid is added up respectively total number of redness, green, black picture element, then divided by the area of this grid, has formed like this color feature vector of 8X8X3=192 dimension.

Based on Gabor gray level image texture feature extraction, we are for the gray level image of the good instrument of normalization for Instrument image, and each gray level image is divided into the grid of 8X8, gets the center of each grid, has generated like this sampled point of 8X8.On each sampled point, we add a Gaussian wave filter, and the Gaussian wave filter here is to try to achieve and come from the Gaussian envelope of Gabor wave filter, and detailed process as the following formula.

$G(x,y)=\frac{{\mathrm{\κ}}^2}{{\mathrm{\σ}}^2}\exp [-\frac{{\mathrm{\κ}}^2(x^2+y^2)}{2{\mathrm{\σ}}^2}]=\frac{4}{{\mathrm{\λ}}^2}\exp [-\frac{2(x^2+y^2)}{{\mathrm{\λ}}^2}]---\left(5\right)$

$F_d(x_i,y_j)=\underset{x=-N}{\overset{x=N}{\mathrm{\Σ}}}\underset{y=-N}{\overset{y=N}{\mathrm{\Σ}}}{f}_d(x_i+x,y_j+y)G(x,y)---\left(6\right)$

Our setting parameter σ=π in the equation above, Wavelength X=8 wherein, N=2 λ.

Like this centered by each sampled point, be that non-zero point is sued for peace by the Gaussian wave filter to pixel values all in the Gaussian envelope, just each sampled point at each template image has obtained an eigenwert, altogether obtains 8X8 (sampled point)=64 dimensional feature vector.Proper vector obtains 64 final dimensional feature vectors by the equation to each eigenwert extraction of square root

3, feature modeling and pattern match

Try to achieve instrument template type Characteristic of Image vector sum and told fortune by analysing the component parts of a Chinese character behind the Characteristic of Image vector of instrument, calculate the Euclidean distance d between them, more show that near 0 this two width of cloth image is more similar.Here establishing test sample book is X, and master sample is that the eigenwert of Y. test sample book is: (x ₁, x ₂, x ₃.., x _n), the eigenwert of master sample is (y ₁, y ₂, y ₃..., y _n), wherein eigenwert all normalizes in the scope of (0.0,1.0).

For color feature value and Gabor feature, itself is the value in (0,1) scope all, and we do not do normalization here.

Discriminant function for each feature is:

$E(x,y)=\underset{k=0}{\overset{n}{\mathrm{\Σ}}}{(x_k-y_k)}^2,k=\mathrm{1,2},...,n---\left(7\right)$

For the selection of sorter, if the special vehicle meter type is fewer, very many of the training sample of every type instrument can adopt neural network classifier to reach reasonable recognition effect.If the type of special vehicle instrument only has 2 classes, a large amount of training samples is arranged, can adopt the sorter of SVM to reach reasonable recognition effect.If the kind of special vehicle meter type is many, thousands of kinds are arranged, can adopt Gaussian modeling or adopt the multi-template matching method to classify and identify, such recognition effect is fast and accurate.

Among the present invention, owing to not being a lot (mainly for 10 kinds of special vehicle meter type for the special vehicle meter type, arrive shown in Figure 12 such as Fig. 3), and we select training sample meticulously to every kind of meter type, the quantity of the training sample of every kind of meter type is not very large (for every kind of meter type, each scale of pointed is got a training sample), all be well-chosen representational, therefore we adopt the arest neighbors rule to carry out type matching, adopt the arest neighbors rule for our this problem extraordinary recognition effect to be arranged.Find the training sample the most contiguous with test sample book (realtime graphic of input) by the most recent method, that is to say the training sample Y of square error minimum, namely:

$Y=\min \left[\underset{k=0}{\overset{n}{\mathrm{\Σ}}}{(x_k-y_k)}^2\right],k=\mathrm{1,2},...,n---\left(8\right)$

If Y belongs to certain meter type M, then the Instrument image of test sample book (realtime graphic of input) also belongs to this meter type M.

4, obtain gauge pointer point of fixity, scale information according to specific meter type

With respect to industrial instrument, the type identification of special vehicle instrument is necessary more, and feasible.At first, the special vehicle meter type is limited, therefore can every kind of special vehicle instrument carry out feature modeling.How a lot of the type of industrial instrument is then, may thousands of kinds.Relative industrial instrument, the dial plate of the instrument that special vehicle is dissimilar takes on a different character, and the feature of industrial instrument is once regular, and the scale of a lot of special vehicle instrument neither be uniform, therefore be necessary very much every kind of dissimilar special vehicle instrument is carried out feature modeling, and every kind of special vehicle instrument is carried out artificial setup parameter value such as its pointer point of fixity, start index and termination scale position, scale respectively.Such as Fig. 3 to Figure 12, showed pointer point of fixity and the scale distribution parameter value of the special vehicle meter type setting that the present invention relates to.After identifying like this type of special vehicle instrument, just can infer important informations such as pointer point of fixity, scale distribution according to the setting value of every kind of special vehicle instrument, if further pointer just can carry out accurate instrument identification after the identification automatically.

The contrast prior art, the present invention has following advantage:

(1) the present invention extracts characteristics of image, and the special vehicle meter type has been carried out modeling and type identification.

(2) color characteristic is for coloured instrument, and is very obvious such as the meter type effect that red color, green color are arranged on the dial plate.

(3) after the present invention has identified meter type automatically, so that the automatic identification of the pointer point of fixity of special vehicle instrument, non-uniform scale all becomes possibility.

5, description of drawings

Fig. 1. based on the automatic identification process figure of the special vehicle meter type of characteristics of image.

Fig. 2. train the modeling process of set based on the special vehicle instrument of characteristics of image.

Fig. 3. determine pointer point of fixity, scale, start index and termination scale according to meter type 1.

Fig. 4. determine pointer point of fixity, scale, start index and termination scale according to meter type 2.

Fig. 5. determine pointer point of fixity, scale, start index and termination scale according to meter type 3.

Fig. 6. determine pointer point of fixity, scale, start index and termination scale according to meter type 4.

Fig. 7. determine pointer point of fixity, scale, start index and termination scale according to meter type 5.

Fig. 8. determine pointer point of fixity, scale, start index and termination scale according to meter type 6.

Fig. 9. determine pointer point of fixity, scale, start index and termination scale according to meter type 7.

Figure 10. determine pointer point of fixity, scale, start index and termination scale according to meter type 8.

Figure 11. determine pointer point of fixity, scale, start index and termination scale according to meter type 9.

Figure 12. determine pointer point of fixity, scale, start index and termination scale according to meter type 10.

Embodiment

Below in conjunction with figure implementation of the present invention is further detailed.

The general flow chart of enforcement of the present invention as shown in Figure 1, its flow process is as follows: after utilizing image capturing system and equipment to obtain a secondary real-time video image of special vehicle instrument, video data is sent in the computing machine, and computing machine carries out automatic identifying processing to the single width Instrument image.At first extract the disk of this width of cloth Instrument image, then according to disc radius Instrument image is carried out normalization, extract respectively color characteristic and Gabor textural characteristics after the normalization, feature templates in these two kinds of features and the training sample is carried out respectively pattern match, then obtain the instrument classification results, at last two kinds recognition result is comprehensively obtained meter type.

Training sample of the present invention obtain feature templates as shown in Figure 2, instrument to every type, meticulously select representational training sample, such as each scale of pointed instrument gets a sample image and is placed in the training set, then extracts color characteristic and Gabor textural characteristics value for each training sample image.

Claims (5)

1. the identification of special vehicle meter type and scaling method based on a characteristics of image is characterized in that, described method is as follows:

At first to every type special vehicle instrument, gather some special vehicle Instrument images, pass through artificial judgment, representational image is stayed be used as training sample; For each training sample, by image pre-processing method picture quality is carried out normalization; Then extract the disk of Instrument image, according to disc radius Instrument image is carried out size normalization; Respectively the special vehicle Instrument image is extracted color characteristic and Gabor textural characteristics after the normalization; Last for all training samples, set up model take color characteristic and Gabor textural characteristics as vector as every type respectively;

After the characteristic model of the training sample database of every type instrument establishes, for the realtime graphic that collects, picture quality normalization is carried out in the pre-service of same process image, extract the instrument disk and carry out the instrument size normalization, extract respectively color characteristic and Gabor textural characteristics after the normalization, feature templates to the instrument training sample of these two kinds of features and each type carries out respectively pattern match, obtains the instrument classification results with the arest neighbors rule.

2. method according to claim 1 is characterized in that, described Instrument image normalization is specific as follows:

At first the special vehicle Instrument image is carried out pre-service, go random noise, stretch to strengthen the contrast of Instrument image by color by medium filtering; Wherein medium filtering adopts the modified adaptive median filter algorithm in conjunction with mean filter and adaptive median filter advantage; And the adaptive neighborhood method of average removal of images that uses band to revise disturbs and noise; Wherein, strengthen the contrast of Instrument image by the color stretching: obtain first gray level image corresponding to coloured image, and the minimum gray value g1 in the statistics Instrument image, gray scale maximal value g2, it is drawn high the gray space of (0,255), so that gray level image is more clear; If g1 be total pixel greater than the minimal gray of n (such as n=10), g2 is that total pixel is greater than the maximum gray scale of n (such as n=10); For any pixel, its gray-scale value is gray (x, y), and the gray-scale value that draws high rear correspondence is GRAY (x, y), then has:

$\mathrm{GRAY}(x,y)=\frac{255\×[\mathrm{gray}(x,y)-g_1]}{(g_2-g_1)}---\left(1\right)$

For the coloured image of RGB pattern, the redness of its correspondence, green, blue color draw high and are respectively:

$R(x,y)=\frac{255\×[r(x,y)-g_1]}{(g_2-g_1)}---\left(2\right)$

$G(x,y)=\frac{255\×[g(x,y)-g_1]}{(g_2-g_1)}---\left(3\right)$

$B(x,y)=\frac{255\×[b(x,y)-g_1]}{(g_2-g_1)}---\left(4\right)$

Secondly, utilize the method for Hough change detection circular arc to obtain the edge of special vehicle instrument disk;

After obtaining the size of instrument disk, the size of instrument is carried out linear normalization, namely crop the outer zone of instrument disk, then the pixel in all instrument disks is normalized into the image that radius is 64 pixel sizes; During linear normalization, only do the conversion of equal proportion in X and Y direction; After the normalization, Instrument image becomes the rectangle of 128x128 size, and the edge of rectangular image is the instrument disk and then, and the pixel in the zone that the instrument disk is outer all is arranged to white pixel.

3. method according to claim 1 is characterized in that, described color characteristic extracts specific as follows:

The Instrument image that normalizing is good is divided into the grid of 8X8, and each grid is added up respectively total number of redness, green, black picture element, then divided by the area of this grid, has formed like this color feature vector of 8X8X3=192 dimension;

Based on Gabor gray level image texture feature extraction, for the gray level image of the good instrument of normalization, each gray level image is divided into the grid of 8X8, gets the center of each grid, has generated like this sampled point of 8X8 for Instrument image; On each sampled point, add a Gaussian wave filter, the Gaussian wave filter here is to try to achieve and come from the Gaussian envelope of Gabor wave filter, detailed process is as the following formula;

$G(x,y)=\frac{{\mathrm{\κ}}^2}{{\mathrm{\σ}}^2}\exp [-\frac{{\mathrm{\κ}}^2(x^2+y^2)}{2{\mathrm{\σ}}^2}]=\frac{4}{{\mathrm{\λ}}^2}\exp [-\frac{2(x^2+y^2)}{{\mathrm{\λ}}^2}]---\left(5\right)$

$F_d(x_i,y_j)=\underset{x=-N}{\overset{x=N}{\mathrm{\Σ}}}\underset{y=-N}{\overset{y=N}{\mathrm{\Σ}}}{f}_d(x_i+x,y_j+y)G(x,y)---\left(6\right)$

Our setting parameter σ=π in the equation above,

Wavelength X=8 wherein, N=2 λ;

Like this centered by each sampled point, be that non-zero point is sued for peace by the Gaussian wave filter to pixel values all in the Gaussian envelope, just each sampled point at each template image has obtained an eigenwert, altogether obtains 8X8 (sampled point)=64 dimensional feature vector; Proper vector obtains 64 final dimensional feature vectors by the equation to each eigenwert extraction of square root

4. method according to claim 1 is characterized in that, described feature modeling and pattern match are specific as follows:

Try to achieve instrument template type Characteristic of Image vector sum and told fortune by analysing the component parts of a Chinese character behind the Characteristic of Image vector of instrument, calculate the Euclidean distance d between them, more show that near 0 this two width of cloth image is more similar; Here establishing test sample book is X, and master sample is that the eigenwert of Y. test sample book is: (x ₁, x ₂, x ₃..., x _n), the eigenwert of master sample is (y ₁, y ₂, y ₃..., y _n), wherein eigenwert all normalizes in the scope of (0.0,1.0);

Discriminant function for each feature is:

$E(x,y)=\underset{k=0}{\overset{n}{\mathrm{\Σ}}}{(x_k-y_k)}^2,k=\mathrm{1,2},...,n---\left(7\right)$

For the selection of sorter, adopt the arest neighbors rule to carry out type matching, find the training sample the most contiguous with test sample book (realtime graphic of input) by the most recent method, that is to say the training sample Y of square error minimum, namely:

$Y=\min \left[\underset{k=0}{\overset{n}{\mathrm{\Σ}}}{(x_k-y_k)}^2\right],k=\mathrm{1,2},...,n---\left(8\right)$

If Y belongs to certain meter type M, then the Instrument image of test sample book (realtime graphic of input) also belongs to this meter type M.

5. method according to claim 1 is characterized in that, described method also further comprises:

After identifying the type of special vehicle instrument, infer according to the setting value of every kind of special vehicle instrument the pointer point of fixity that the important information that scale distributes further carries out accurate instrument identification.

Images