WO2020132960A1

WO2020132960A1 - Defect detection method and defect detection system

Info

Publication number: WO2020132960A1
Application number: PCT/CN2018/123966
Authority: WO
Inventors: 王星泽; 舒远
Original assignee: 合刃科技（深圳）有限公司
Priority date: 2018-12-26
Filing date: 2018-12-26
Publication date: 2020-07-02
Also published as: CN111344559B; CN111344559A

Abstract

A defect detection method, comprising: using coherent light sources (11, 12, 13, 14) of different wavelengths to illuminate an object to be inspected; capturing multiple speckle images generated by said object under the illumination of the coherent light sources (11, 12, 13, 14) of different wavelengths; and detecting a defect of said object by means of the speckle images. The beneficial effects are that: the measurement accuracy reaches the level of optical wavelength and micro-defects at a micron level can be detected; a non-contact, high-accuracy, online, and real-time non-destructive detection method is achieved.

Description

Defect detection method and defect detection system

Technical field

The invention relates to the technical field of inspection, in particular to a defect detection method and a defect detection system.

Background technique

In the prior art, metal products will leave some on the surface of metal products due to friction, cutting deformation and tearing and processing environment changes during mechanical processing, chemical processing, spray coating, etc. The bad defects with different shapes and sizes mainly include small defects such as shrinkage holes, bubbles, cracks, peeling, white spots, and intergranular cracks that cannot be recognized by the naked eye. At the same time, for curved objects with low texture and high reflectivity, such as packaging tin balls (BGA tin balls), high-brightness metal balls, mobile phone metal shells, etc., their surface texture features are single or even missing, and their surfaces are smooth and extremely strong Reflective characteristics, causing it to easily produce too bright light spots.

A conventional detection method in the prior art is the Automatic Optical Inspection (AOI) method. The AOI detection method performs direct or indirect microscopic magnification on the detection object, and uses digital image algorithms to target after microscopic imaging. Segmentation recognition process to detect various defect areas on the product surface. The AOI detection method requires high vertical and horizontal resolution of the optical detection system, and the high reflection and surface curvature of the metal product surface will cause uneven illumination, that is, the illumination light source has a great influence on the detection result in the AOI detection. Large, the illumination light source must be suitable for the detection of various defects, and the detection of various defects can perform well without losing any defect information.

Another conventional detection method in the prior art is a three-dimensional reconstruction method based on active structured light projection. However, due to the large area of flares, it will affect the extraction of grating stripes, resulting in the inability to obtain accurate depth information in the detection. Large areas of data holes, although spraying white powder on the surface of metal products can reduce the occurrence of such errors, but these white powder will block the defects and make the detection system lose its detection ability.

The inventor found through research that both the two-dimensional imaging method and the optical three-dimensional scanning method commonly used in the prior art are difficult to perform effective defect detection on metal surfaces; meanwhile, in the prior art, most of the light sources used in traditional detection systems are incoherent Light source, its detection sensitivity is relatively low compared to the detection system using coherent light source, especially for micron-scale scratch defects or pitted surface defects, incoherent light is incident on the side surface in parallel and reflected in all directions, causing defects The imaging contrast with the background area is not large, and the defect will be completely submerged in the background interference and cannot be detected.

Summary of the invention

Based on this, in order to solve the technical problems in the prior art, a defect detection method is specifically proposed, including:

Use coherent light sources of different wavelengths to illuminate the measured object;

Shooting the object under test under the illumination of the coherent light sources of different wavelengths to generate a speckle image;

Detecting defects of the measured object using a neural network and the speckle image, inputting the speckle image and/or feature parameters extracted from the speckle image to the neural network, and outputting the neural network A detection result, where the detection result is a defect type of the object to be tested.

In one embodiment, the feature parameters extracted from the speckle image include the speckle extension rate calculated by the autocorrelation function of the speckle image.

In an embodiment, the feature parameters extracted from the speckle image include first-order statistical characteristics and second-order statistical characteristics of the speckle image obtained by calculation.

In one embodiment, the use of a neural network and the speckle image to detect defects of the measured object includes a training phase and a detection phase;

The training phase includes:

Classify the different defects of multiple test object samples, and use the classification results as the output characteristics of the output layer of the neural network; switch different wavelengths to capture multiple speckle images of the multiple test object samples, and collect more for each type of defect A speckle image on the surface of the measured object, using the speckle image and/or the feature parameters extracted from the speckle image to form a training data set for training a neural network;

Use the above speckle image that can indirectly reflect the fine structure of the surface and/or the feature parameters extracted from the speckle image as the input features of the input layer of the neural network, and use the classification results as the output features of the output layer of the neural network. The neural network is trained by the features and output features, and the neural network model of the relationship between the speckle image on the surface of the measured object and/or the feature parameters extracted from the speckle image and the surface defect of the measured object is obtained by training;

The detection phase includes:

Input speckle images corresponding to the coherent light sources of different wavelengths and/or feature parameters extracted from the speckle image into the input layer of the trained neural network model for detection, wherein the speckle image and/ Or the feature parameter extracted from the speckle image is the input feature of the neural network;

The output layer of the neural network outputs a detection result, and the detection result includes one or more of defect-free, air bubble, and deformation.

In one embodiment, before using the neural network and the speckle image to detect the defect of the measured object, one of the methods including neighborhood mean filtering, median filtering, low-pass filtering, and homomorphic filtering or A variety of filters for noise in the speckle image.

In addition, in order to solve the technical problems in the prior art, a defect detection system is specifically proposed, including:

The defect detection system includes a coherent light source group, a light source controller, a beam adjustment module, a photoelectric sensor module, and a detection module;

Wherein, coherent light sources of different wavelengths constitute the coherent light source group of the defect detection system;

Wherein, the light source controller uses a software program switch to control the coherent light source group to achieve switching between coherent light sources of different wavelengths;

Wherein, the beam adjustment module includes a beam combiner, a collimated beam expander, and a beam splitter; the beam adjustment module is used to adjust the optical path formed by the coherent light source to form a speckle image on the photoelectric sensor module;

Wherein, the photoelectric sensing module includes one or more photoelectric sensors, and the speckle image is captured by the photoelectric sensor, and the captured speckle image is transmitted to the detection module for detection processing;

Wherein, the detection module uses a neural network and the speckle image to detect defects of the measured object, and inputs the speckle image and/or feature parameters extracted from the speckle image to the neural network, which is used by The neural network outputs a detection result, and the detection result is a defect type of the detected object.

In an embodiment, the adjusting the optical path formed by the coherent light source using the beam adjustment module to form a speckle image on the photoelectric sensor module includes that the coherent light source is a laser;

Use the beam combiner to superimpose the laser beams of different wavelength laser light sources at the same height and at the same time into one beam;

The laser beam combined by the beam combiner then passes through a collimating beam expander, which makes the outgoing laser beam project into a spot with uniform intensity distribution on the white screen, and the outgoing laser beam is parallel Laser beam, use collimating beam expander to complete the beam collimation process;

The parallel laser beam is deflected after passing through the reflecting mirror, and then irradiated to the surface of the measured object through the beam splitter. The scattered light on the surface of the measured object is reflected by the beam splitter and enters the photoelectric sensor.

In one embodiment, the photoelectric sensing module includes one or more imaging lenses, and the scattered light reflected from the surface of the measured object first passes through the imaging lens and then is projected into the photoelectric sensor;

In another embodiment, the photoelectric sensor module does not include an imaging lens, and the scattered light reflected from the surface of the measured object is directly projected onto the photoelectric sensor; the above-mentioned non-lens imaging method can avoid when the distance of the measured object The change of the focal plane is necessary for the imaging to be clear when changing, so it is suitable for detecting objects with a curved surface.

In an embodiment, the defect detection system uses a liquid crystal tunable filter to switch between different wavelengths, and the control electrical signal sent by the light source controller transforms the filter band of the liquid crystal tunable filter.

In one embodiment, the detection module uses a neural network and the speckle image to detect the defect of the measured object, including a training phase and a detection phase;

The training phase includes:

The detection phase includes:

In one embodiment, the photoelectric sensor is a CCD photoelectric sensor or a CMOS photoelectric sensor.

The implementation of the embodiments of the present invention will have the following beneficial effects:

The detection method disclosed in the present invention has a detection accuracy that reaches the order of light wave wavelength and can detect micro-level (um) micro-defects. It is a non-contact, high-precision, online, and real-time non-destructive detection method.

BRIEF DESCRIPTION

The drawings needed to be used in the embodiments or the description of the prior art will be briefly introduced below.

In order to more clearly explain the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only These are some embodiments of the present application. For those of ordinary skill in the art, without paying any creative labor, other drawings can be obtained based on these drawings.

1 is a schematic diagram of a defect detection system in the present invention;

2 is a schematic diagram of a defect detection method based on a neural network in the present invention.

detailed description

The technical solutions in the embodiments of the present invention will be described clearly and completely in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by a person of ordinary skill in the art without making creative efforts fall within the protection scope of the present invention.

After the coherent light source is reflected by the surface, an interference pattern is formed. By analyzing the spatial distribution of the interference pattern, the presence of micro defects can be effectively detected. In the technical solution of the present invention, the multi-speckle spreading effect is used to detect defects on the surface of metal products. After the coherent light source illuminates the surface of the metal product, the amplitude and phase of the outgoing light field function carry a lot of information on the microstructure of the metal product surface; after the combination of coherent light sources of different wavelengths, they can be incident on the surface of the measured object coaxially and time-sharing. The diameter of the speckle is proportional to the wavelength of the coherent light source, the monochromatic speckles of different wavelengths are misaligned with each other in the radial direction, and the multicolor speckle field synthesized by the monochromatic speckles will have speckle elongation in a circular area . The extent of speckle speckle extension depends on the microstructure of the surface of the metal product being tested: for smooth and flat surfaces, due to small spatial changes in the microstructure scale, small changes in the wavelength of the coherent light emitted by the coherent light source will cause the speckle to change drastically, thus Makes the speckle plaque prolong by a larger extent; and for those surface areas with defects, the microstructure of the surface area changes are in the micrometer (um) or millimeter (mm) level, which changes the wavelength of the coherent light emitted by the coherent light source Insensitive, the spread of speckles and plaques is smaller. Thus, the intensity distribution signal of the multi-dispersive speckle field is collected by the photoelectric imaging device, and the speckle elongation rate can be obtained through autocorrelation calculation, thereby detecting fine defects on the metal surface, such as scratches, pits, wear points and other defects.

At the same time, the defect detection method in the technical solution of the present invention is also applicable to the measured objects of some specific materials, such as objects made of translucent plastic materials, objects mixed with different materials, and surface defects such as shallow bubbles and pores. In the detection method, when the coherent light source interferes, the light projected by the light source will penetrate into the object, so that the internal information of the object will also be displayed in the speckle signal and recognized. The technical solution of the present invention can be used for defect detection in various other fields, such as changes in product shape, changes in internal material structure of products, physical damage of products, changes in product colors, etc.

The invention discloses a defect detection system. As shown in FIG. 1, it includes a coherent light source group 1, a light source controller 2, a beam adjustment module 3, a photoelectric sensor module 4, and a detection module (not shown in the figure);

1) Coherent light source group 1, because the laser has the advantages of good monochromaticity, good linearity, and stable output, using lasers of different wavelengths as the coherent light source of the detection system;

In one embodiment, multiple groups of lasers of different wavelengths are used as coherent light sources of the detection system to form a coherent light source group; wherein, the multiple groups of laser light sources of different wavelengths may be any of 2, 3, or 4 groups Configured laser light sources; preferably, as shown in FIG. ₁ , four groups of

laser light sources

11, 12, 13, 14 with wavelengths of λ ₁ , λ ₂ , λ ₃ , and λ ₄ are used as coherent light source groups;

In another embodiment, the detection system uses a liquid crystal tunable filter to switch between different wavelengths, the liquid crystal tunable filter is fixed in front of the photoelectric sensor, and a control electrical signal sent by the light source controller Quickly change the filter band of the liquid crystal tunable filter, which can realize the wavelength selection of at least 10nm, thereby realizing more accurate detection of surface microstructure;

2) Light source controller 2, the light source controller uses software program switches to control the coherent light source group to achieve switching between laser light sources of different wavelengths;

3) The beam adjustment module 3 includes a beam combiner 31, a collimated beam expander 32, a mirror 33, and a beam splitter 34; since multiple laser beams in the detection system need to be coaxial, the beam combiner 31 is used When the laser light sources are emitted at the same height and at the same time, the laser beams coincide into one beam; the laser beams combined by the beam combiner 31 then pass through the collimated beam expander 32, and the collimated beam expander 32 makes the emitted laser beam Projected on the white screen as a spot with uniform light intensity distribution, and the emitted laser beam is a parallel laser beam, using a single large aperture lens in the collimating beam expander 32 to complete the beam collimation process; then, the parallel laser The light beam is deflected after passing through the reflecting mirror 33, and then irradiated to the surface of the object to be measured by the beam splitter 34 with a split ratio of 1 _: 1, wherein the scattered light on the surface of the object to be tested is reflected by the beam splitter and enters the photoelectric sensor 42.

4) The photoelectric sensor module 4 includes one or more photoelectric sensors 42 and an imaging lens 41. The photoelectric sensors 42 are used to photograph speckles; the captured speckle images are transmitted to the detection module for detection processing; Wherein, the imaging lens 41 can be set or removed according to requirements;

In one embodiment, the photoelectric sensor 42 is a CCD photoelectric sensor or a CMOS photoelectric sensor;

In one embodiment, the scattered light reflected back from the surface of the measured object first passes through the imaging lens 41 and then is projected into the photoelectric sensor 42 for digital imaging processing;

In another embodiment, the imaging lens 41 is removed, that is, the photoelectric sensor module 4 does not need the imaging lens 41, and directly reflects the scattered light reflected from the surface of the measured object onto the photoelectric sensor 42; A non-lens imaging method can avoid the problem of adjusting the focal plane for clear imaging when the distance of the measured object changes. Therefore, it is suitable for detecting objects with a curved surface.

In one embodiment, one or more photoelectric sensors are used to detect the speckle images obtained from different angles, and further detection and recognition processing is performed; this method is suitable for non-lens imaging methods for detection and also for lens imaging Measurement.

5) A detection module, which uses a neural network and the speckle image to detect defects of the measured object, and inputs the speckle image and/or feature parameters extracted from the speckle image to the neural network , The detection result is output by the neural network, and the detection result is a defect type of the detected object.

The invention discloses a surface defect detection method, including:

In one embodiment, a plurality of laser light sources in the coherent light source group as multi-color incident laser light sources are respectively switched on and off by the light source controller, and multiple monochromatic speckle images corresponding to lasers of different wavelengths are respectively captured;

In an embodiment, the plurality of different wavelengths are respectively four wavelengths λ ₁ , λ ₂ , λ ₃ , and λ ₄ ;

In one embodiment, the speckle lengthening rate calculated by the speckle autocorrelation function is used to detect the surface defect of the measured object.

The effect of multicolor speckle extension is closely related to the wavelength combination of the incident laser light source; where the wavelength difference determines the position of the speckle extension region; for a fixed combination of laser wavelength changes, on a smooth, defect-free object surface, the object surface The roughness and the wavelength scale of the incident laser light source are both in the nanometer (nm) level. At this time, the speckle particle extension due to the change of the wavelength of the laser light source is larger; and for the surface of the object with the concave and convex defects, the physical scale of the defect is Micron (um) and millimeter (mm) levels are insensitive to changes in the wavelength of the light source, and the speckle particle extension is small; thus, the speckle extension rate calculated by the speckle autocorrelation function can effectively detect the microscopic surface of the object defect.

In one embodiment, the speckle image is regarded as a texture image, and the first-order statistical characteristic and the second-order statistical characteristic are calculated as the texture by calculating the first-order statistical characteristic and the second-order statistical characteristic of the speckle image Feature parameters, establish the correlation model between the normal surface and the surface of the defect area and the texture features of the speckle image, use the significantly associated texture feature parameters to characterize different defects; use multiple speckle images obtained by shooting the object under test to calculate For the texture feature parameter, the correlation model is used to detect the defect corresponding to the texture feature.

The multicolor laser dynamic speckle image can provide a lot of useful information. The dynamic characteristics of the laser speckle image can be regarded as a texture image by means of statistical methods. By calculating the first-order statistical characteristics of the texture image and the second First-order statistical characteristics, the texture feature correspondence model of the normal surface and the defective area surface and the dynamic speckle image is obtained, and the texture feature parameters that are significantly associated are found to detect defects.

In one embodiment, since there is a large amount of random noise in the speckle image, eliminating or reducing the random noise is a prerequisite for accurately obtaining the speckle image information; it may include neighborhood mean filtering, median filtering, low-pass filtering, One or more methods in homomorphic filtering to filter out noise.

In an embodiment of the present invention, as shown in FIG. 2, a neural network is used for learning training of speckle data with large samples and different wavelengths, and the neural network is used for defect detection; the neural network is based on deep learning Deep neural network.

The surface defect detection method includes a training phase and a detection phase;

During the training phase:

Classify different defects of multiple test object samples, and use the classification results as the output characteristics of the neural network output layer; switch different wavelengths to capture multiple speckle images of the multiple test object samples, and collect 1000 for each type of defect Speckle images of the measured object surface and above constitute a training data set for training neural networks;

Among them, the output features include no defects, bubbles, deformation, etc.;

Using the above speckle image that can indirectly reflect the fine structure of the surface as the input features of the input layer of the neural network, and the classification results as the output features of the output layer of the neural network, the neural network is trained using the input features and output features to obtain the A neural network model of the relationship between the speckle image on the surface of the measured object and the surface defects of the measured object.

During the inspection phase:

Through the light source controller, the multiple groups of laser light sources that are multi-color incident laser light sources in the coherent light source group are respectively switched on and off, and the single-color speckle images corresponding to multiple lasers of different wavelengths are captured respectively;

Input the speckle image signal corresponding to the multi-color laser into the input layer of the neural network for detection, wherein the speckle image signal is an input feature of the neural network;

The output layer of the neural network outputs a detection result, and the detection result includes output features such as no defects, bubbles, and deformation.

Many industries need to detect micro-defects on metal surfaces, and detect micro-defects on smooth metal surfaces, such as scratches, pits, and wear points. The illumination optical design and photoelectric sensor adopted in the present invention can be modified to adapt to different application scenarios, so as to realize the requirements of different detection resolutions. The measurement accuracy of the invention reaches the order of light wave wavelength, and can detect micro defects in the micrometer (um) level. It is a non-contact, high-precision, online, and real-time non-destructive detection method.

The above disclosure is only the preferred embodiments of the present invention, and of course cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention still fall within the scope of the present invention.

Claims

A defect detection method, which includes:

Use coherent light sources of different wavelengths to illuminate the measured object;

Taking a speckle image generated by the object under test under the illumination of the coherent light sources of different wavelengths;

Detecting defects of the measured object using a neural network and the speckle image, inputting the speckle image and/or feature parameters extracted from the speckle image to the neural network, and outputting the neural network A detection result, where the detection result is a defect type of the object to be tested.
The defect detection method according to claim 1, wherein

The feature parameters extracted from the speckle image include the speckle extension rate calculated by the autocorrelation function of the speckle image.
The defect detection method according to claim 1, wherein

The feature parameters extracted from the speckle image include first-order statistical characteristics and second-order statistical characteristics of the speckle image obtained by calculation.
The defect detection method according to any one of claims 1-3, characterized in that

The use of a neural network and the speckle image to detect defects of the measured object includes a training phase and a detection phase;

The training phase includes:

Classify different defects of multiple test object samples, and use the classification results as the output characteristics of the output layer of the neural network; switch different wavelengths to capture multiple speckle images of the multiple test object samples, and collect more for each type of defect A speckle image on the surface of the measured object, using the speckle image and/or feature parameters extracted from the speckle image to form a training data set for training a neural network;

Use the speckle image and/or feature parameters extracted from the speckle image as input features of the neural network input layer, and use the classification results as output features of the neural network output layer, and use the input features and output features to neural The network is trained to obtain a neural network model of the relationship between the speckle image on the surface of the measured object and/or the characteristic parameters extracted from the speckle image and the surface defect of the measured object;

The detection phase includes:

Input speckle images corresponding to the coherent light sources of different wavelengths and/or feature parameters extracted from the speckle image into the input layer of the trained neural network model for detection, wherein the speckle image and/ Or the feature parameter extracted from the speckle image is the input feature of the neural network;

The output layer of the neural network outputs a detection result, and the detection result includes one or more of defect-free, air bubble, and deformation.
The defect detection method according to claim 4, characterized in that

Before detecting the defects of the measured object using the neural network and the speckle image, one or more of neighborhood mean filtering, median filtering, low-pass filtering, and homomorphic filtering are used to filter out the scattered Speckle noise.
A defect detection system, characterized in that the defect detection system includes a coherent light source group, a light source controller, a beam adjustment module, a photoelectric sensor module and a detection module;

Wherein, coherent light sources of different wavelengths constitute the coherent light source group of the defect detection system;

Wherein, the light source controller uses a software program switch to control the coherent light source group to achieve switching between coherent light sources of different wavelengths;

Wherein, the beam adjustment module includes a beam combiner, a collimated beam expander, and a beam splitter; the beam adjustment module is used to adjust the optical path formed by the coherent light source to form a speckle image on the photoelectric sensor module;

Wherein, the photoelectric sensing module includes one or more photoelectric sensors, and the speckle image is captured by the photoelectric sensor, and the captured speckle image is transmitted to the detection module for detection processing;

Wherein, the detection module uses a neural network and the speckle image to detect defects of the measured object, and inputs the speckle image and/or feature parameters extracted from the speckle image to the neural network, which is used by The neural network outputs a detection result, and the detection result is a defect type of the detected object.
The defect detection system according to claim 6, wherein:

The adjusting the optical path formed by the coherent light source using the beam adjustment module to form a speckle image on the photoelectric sensor module includes that the coherent light source is a laser;

First, use the beam combiner to superimpose the laser beams of different wavelength laser light sources at the same height and at the same time into one beam;

Secondly, the laser beam combined by the beam combiner then passes through a collimating beam expander, which makes the outgoing laser beam project on the white screen into a spot with uniform intensity distribution, and the outgoing laser beam is For collimated laser beams, use a collimating beam expander to complete the beam collimation process;

Finally, the parallel laser beam is deflected after passing through the reflecting mirror, and then irradiated to the surface of the measured object through the beam splitter. The scattered light on the surface of the measured object is reflected by the beam splitter and enters the photoelectric sensor.
The defect detection system according to claim 7, characterized in that

Wherein, the photoelectric sensing module includes one or more imaging lenses, and the scattered light reflected from the surface of the measured object first passes through the imaging lens and then is projected into the photoelectric sensor.
The defect detection system according to claim 7, characterized in that

Wherein, the photoelectric sensing module does not include an imaging lens, and the scattered light reflected from the surface of the measured object is directly projected onto the photoelectric sensor.
The defect detection system according to any one of claims 6-9, characterized in that

The defect detection system uses a liquid crystal tunable filter to switch between different wavelengths, and transforms the filter band of the liquid crystal tunable filter through a control electrical signal sent by a light source controller.