CN111223093A - AOI defect detection method - Google Patents
AOI defect detection method Download PDFInfo
- Publication number
- CN111223093A CN111223093A CN202010141504.0A CN202010141504A CN111223093A CN 111223093 A CN111223093 A CN 111223093A CN 202010141504 A CN202010141504 A CN 202010141504A CN 111223093 A CN111223093 A CN 111223093A
- Authority
- CN
- China
- Prior art keywords
- image
- defect detection
- defect
- network
- abnormal
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000007547 defect Effects 0.000 title claims abstract description 199
- 238000001514 detection method Methods 0.000 title claims abstract description 143
- 230000002159 abnormal effect Effects 0.000 claims abstract description 81
- 238000012549 training Methods 0.000 claims abstract description 53
- 238000012545 processing Methods 0.000 claims description 31
- 238000000034 method Methods 0.000 claims description 30
- 238000005070 sampling Methods 0.000 claims description 22
- 238000000605 extraction Methods 0.000 claims description 21
- 230000008569 process Effects 0.000 claims description 21
- 238000010606 normalization Methods 0.000 claims description 11
- 238000012360 testing method Methods 0.000 claims description 10
- 238000007689 inspection Methods 0.000 claims description 6
- 238000001914 filtration Methods 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- 238000010200 validation analysis Methods 0.000 claims 1
- 230000003287 optical effect Effects 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 12
- 238000010586 diagram Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 6
- 230000011218 segmentation Effects 0.000 description 5
- 230000000007 visual effect Effects 0.000 description 4
- 230000005856 abnormality Effects 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 230000002950 deficient Effects 0.000 description 3
- 230000000877 morphologic effect Effects 0.000 description 3
- 239000013598 vector Substances 0.000 description 3
- 238000012795 verification Methods 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000011478 gradient descent method Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000008485 antagonism Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000010923 batch production Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000003708 edge detection Methods 0.000 description 1
- 230000008014 freezing Effects 0.000 description 1
- 238000007710 freezing Methods 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 230000036039 immunity Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 230000001902 propagating effect Effects 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20092—Interactive image processing based on input by user
- G06T2207/20104—Interactive definition of region of interest [ROI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
- G06T2207/30121—CRT, LCD or plasma display
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the technical field of automatic optical detection of display panels, and discloses an AOI defect detection method, which comprises the steps of establishing a defect detection model, wherein the defect detection model comprises a generator network and a discriminator network and is used for acquiring defect related information; establishing a training set by using the positive sample, and training a generator network and a discriminator network by using the training set; updating the defect detection model by utilizing the generator network and the discriminator network obtained by training to obtain an updated defect detection model; collecting an image of a panel to be detected; and inputting the image of the panel to be detected into the updated defect detection model, detecting the defect and obtaining defect detection information. The invention solves the problems of obvious abnormal missing detection areas, insufficient utilization of a large number of positive samples and few negative samples in the panel defect detection method based on supervised learning in the prior art, can improve the defect detection capability of the whole AOI system and reduce the detection cost.
Description
Technical Field
The invention relates to the technical field of automatic optical detection of display panels, in particular to an AOI defect detection method.
Background
The automatic optical inspection system (AOI) is widely applied in industry, can replace people to finish a series of work with high repeatability and certain risk, has the characteristics of high speed, high precision, good reliability, no contact, no damage, high cost performance, easy function expansion and the like, and greatly improves the generation efficiency.
A panel defect detection system is arranged in an AOI system in a display panel manufacturing line, and the detection accuracy of the system is extremely important to ensure the quality of panels manufactured by manufacturers. For the defect detection of the display panel, no matter appearance detection or module point screen detection, firstly, a high-precision camera carries out high-quality multi-view-angle photographing on the panel, then, the detection algorithm of the AOI system is utilized to detect and calibrate the defects possibly existing in the panel, and corresponding processing is carried out on different defects. The original image shot by the high-precision camera has very high resolution, typically 6000x15000, but the defects on the display panel are different in shape, some defects are obvious (such as appearance defects), and some defects are invisible to naked eyes, such as bright and dark spot defects and some MURA defects. The primary task of the AOI defect detection system of the LCD/OLED is to locate possible defects in a high-resolution image obtained from a high-precision camera, find the problems of the panel in time and further improve the yield of a production line. Because the resolution of the input image is very high and the defects are relatively very small, the method puts high requirements on the defect detection visual algorithm.
At present, visual algorithms for defect detection in a display panel AOI detection system are mainly based on two types, one is a traditional computer visual algorithm based on rules and logic, and the other is a machine learning algorithm based on sample data, wherein the deep learning visual algorithm mainly takes CNN as a core.
For a traditional computer vision algorithm, defect detection is mainly performed based on a manual feature extraction mode, for example, a traditional defect segmentation algorithm based on hierarchical feature extraction, threshold transformation and mathematical morphology. For the algorithm, the detection result has the characteristic of strong interpretability due to the fact that the theoretical basis is sufficient. However, in the algorithm based on the traditional computer vision, because different change conditions of the defects in a possible high-dimensional space need to be considered, the adjustment hyper-parameters of the algorithm are correspondingly increased, meanwhile, the traditional algorithm is poor in robustness and cannot be well adapted to the change of defect forms, the effect of manually extracted features is unstable, and the effect of the algorithm is further reduced.
As described above, the conventional algorithm has poor feature extraction stability and needs more adjustment parameters, so the display panel defect detection is gradually shifted to the deep learning algorithm. Also, deep neural networks are different according to different requirements in the field. For some sites, firstly, a traditional computer vision sliding window matching algorithm is used for roughly positioning the defects, and then, a classification network is used for accurately classifying the defects; for the fields with high requirements, the detection network is directly used for accurate positioning of the defects, and for the fields with higher requirements, the segmentation network is also used for accurate segmentation of the defects.
For all the classification detection segmentation networks, learning is performed based on display panel defect samples, and all samples are known defects and are labeled manually, so that most of the current panel defect detection algorithms based on deep learning are supervised learning. Industrial practice shows that the defect patterns on display panels vary widely, a defect sample cannot cover all kinds of defects at all, old defects are artificially classified into new ones, and new unknown defects (hereinafter referred to as abnormal defects) are sequentially generated. Because supervised learning is adopted, the category must be constant during training, so the CNN can only automatically extract the category features in the training set, and for abnormal defects occurring in the actual field, the CNN can miss detection of such defects (abnormal defects) in most cases because the abnormal defects do not occur in the training set, and the abnormal defects are generally obvious large-scale obvious abnormalities, and customers cannot accept the obvious abnormal missing detection.
A further disadvantage of supervised learning based defect detection is that the data sets on which the training set depends are all negative examples, i.e. known defects that are manually labeled. However, in the field of industrial defect detection of display panels, the yield of the panel industry is very high, i.e. there are a large number of positive samples and very few negative samples in the sample set. The supervised learning-based method just discards the most positive samples, and only uses a small amount of negative samples to learn. Because the number of samples is very small, data expansion to a certain degree is often carried out, the data expansion cannot solve the fundamental problem of few samples, the expanded data can only slightly improve the accuracy of defect detection, and the problem that negative samples segmented in a specific scene of a display panel by supervised classification detection are sparse in a high-dimensional space cannot be fundamentally solved.
Disclosure of Invention
The embodiment of the application provides an AOI defect detection method, and solves the problems that in the prior art, a panel defect detection method based on supervised learning has obvious missed detection defects, a large number of positive samples are not fully utilized, and negative samples are few.
The embodiment of the application provides an AOI defect detection method, which comprises the following steps:
establishing a defect detection model, wherein the defect detection model comprises a generator network and a discriminator network, and is used for obtaining defect related information, and the defect related information comprises an abnormal score of an input sample and/or specific position information of an abnormal area;
establishing a training set by using positive samples, and training the generator network and the discriminator network by using the training set;
updating the defect detection model by using the generator network and the discriminator network obtained by training to obtain an updated defect detection model;
collecting an image of a panel to be detected;
and inputting the image of the panel to be detected into the updated defect detection model, detecting the defects, and obtaining the related information of the defects of the panel to be detected.
Preferably, the specific location information of the abnormal region includes an abnormal region outline and/or an abnormal region bounding box.
Preferably, the generator network adopts an Unet structure, and the input image is subjected to N-layer (N > = 1) down-sampling and N-layer up-sampling to obtain a reconstructed image;
the down-sampling comprises three sub-processing processes, namely a LeakyRelu layer, a convolution layer and a batch normalization layer in sequence; the up-sampling comprises three sub-processing processes, namely a Relu layer, a transposed convolution layer and a batch normalization layer.
Preferably, the input image of the discriminator network is subjected to N (N > = 1) layer down-sampling, feature extraction, and convolution processing to obtain an output score;
the down-sampling and the feature extraction both comprise three sub-processing processes, namely a LeakyRelu layer, a convolution layer and a batch normalization layer in sequence; the convolution processing comprises two sub-processing processes, namely a convolution layer and a sigmoid layer.
Preferably, the defect detection model comprises an abnormal GAN forward module and a feature information extraction module which are connected in sequence;
the abnormal GAN forward module comprises a generator network and a discriminator network which are connected in sequence and is used for obtaining the abnormal score and the residual map of the input sample;
the characteristic information extraction module comprises a residual error map opening and closing operation module, an abnormal region outline extraction module and an outline filtering module which are sequentially connected and is used for obtaining an abnormal region outline and an abnormal region boundary frame according to the residual error map.
Preferably, the training of the generator network and the discriminator network by using a training set is:
for each positive sample, the real image is used as an input image, and a reconstructed image is generated after the input image passes through a generator network; inputting the reconstructed image into a discriminator network, and outputting a reconstruction score and a reconstruction characteristic; the real image is used as an input image, and a real score and a real feature are output after the real image directly passes through the discriminator network;
obtaining a minimized generator loss function and a discriminator loss function from the real image, the reconstructed image, the real feature, the reconstructed feature, the real score, and the reconstructed score.
Preferably, the generator loss function comprises residual loss, countermeasure loss, and feature loss; the discriminator loss function comprises real graph confrontation loss and reconstructed graph confrontation loss;
and training the corresponding network by back propagation alternately according to the generator loss function and the discriminator loss function by utilizing the training set.
Preferably, after each round of training is finished, testing is carried out through a verification set, and an AUC value is calculated to serve as a test index of the round of training; and after all training is finished, storing the generator network and the discriminator network corresponding to the highest AUC value, and using the generator network and the discriminator network in the updated defect detection model.
Preferably, before the inputting the panel image to be detected into the updated defect detection model, the method further includes: preprocessing the panel image to be detected;
the pretreatment comprises the following steps: extracting an ROI (region of interest) region of the panel image to be detected to obtain an ROI region map; segmenting the ROI regional graph to obtain a plurality of first images; segmenting each first image to obtain a plurality of second images serving as batch processing images;
and inputting the batch processing images into the updated defect detection model to obtain defect related information corresponding to each second image.
Preferably, the AOI defect detection method further includes performing defect classification, specifically:
obtaining known defect information according to the panel image to be detected; and judging whether the defect is a known defect or an abnormal defect by using the overlapping area according to the defect related information and the known defect information.
One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:
in the embodiment of the application, training is performed by discarding a small number of labeled negative samples, and only positive samples with overwhelming number are used for training, and the training mode is based on GAN, so that the method is an unsupervised training mode, a GAN network is trained on a large number of positive samples to only learn the characteristics of the normal samples and the manifold of a high-dimensional space, GAN can obtain defect related information (abnormal score, abnormal region outline and abnormal region boundary frame of an input sample), defect detection information can be obtained after a panel image to be detected is input, the normal samples and the defect samples are effectively distinguished, the defects can be effectively distinguished into known defects or abnormal defects (abnormal samples which are not generated at all) by combining the known defect information, therefore, missing detection of the defects, particularly abnormal defects, on the actual site can be effectively avoided, and customer satisfaction is improved, further improving the defect detection capability of the whole AOI system.
Drawings
In order to more clearly illustrate the technical solution in the present embodiment, the drawings needed to be used in the description of the embodiment will be briefly introduced below, and it is obvious that the drawings in the following description are one embodiment of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.
Fig. 1 is a schematic diagram of a generator network in an AOI defect detection method according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a discriminator network in an AOI defect detection method according to an embodiment of the present invention;
FIG. 3 is a general structural diagram of GAN resistance training in the AOI defect detection method according to an embodiment of the present invention;
FIG. 4 is a general flowchart of an AOI defect detection method according to an embodiment of the present invention;
FIG. 5 is a flowchart illustrating an anomaly detection method in an AOI defect detection method according to an embodiment of the present invention;
fig. 6 is a sample diagram of the results of performing anomaly detection on the appearance of an LCD display panel by using an AOI defect detection method according to an embodiment of the present invention.
Detailed Description
In order to better understand the technical solution, the technical solution will be described in detail with reference to the drawings and the specific embodiments.
The embodiment provides an AOI defect detection method, which comprises the following steps:
step 1, establishing a defect detection model, wherein the defect detection model comprises a generator network and a discriminator network, and is used for obtaining defect related information.
Wherein the defect-related information comprises an abnormal score and/or an abnormal region outline and/or an abnormal region bounding box of the input sample.
Specifically, the defect detection model comprises an abnormal GAN forward module and a characteristic information extraction module which are connected in sequence; the abnormal GAN forward module comprises a generator network and a discriminator network which are connected in sequence and is used for obtaining the abnormal score and the residual map of the input sample; the characteristic information extraction module comprises a residual error map opening and closing operation module, an abnormal region outline extraction module and an outline filtering module which are sequentially connected and is used for obtaining an abnormal region outline and an abnormal region boundary frame according to the residual error map.
The generator network adopts an Unet structure, and an input image is subjected to N-layer (N > = 1) down-sampling and N-layer up-sampling processing to obtain a reconstructed image; the down-sampling comprises three sub-processing processes, namely a LeakyRelu layer, a convolution layer and a batch normalization layer in sequence; the up-sampling comprises three sub-processing processes, namely a Relu layer, a transposed convolution layer and a batch normalization layer.
The input image of the discriminator network is subjected to N (N > = 1) layers of downsampling, feature extraction and convolution processing to obtain an output score; the down-sampling and the feature extraction both comprise three sub-processing processes, namely a LeakyRelu layer, a convolution layer and a batch normalization layer in sequence; the convolution processing comprises two sub-processing processes, namely a convolution layer and a sigmoid layer.
And 2, establishing a training set by using the positive sample, and training the generator network and the discriminator network by using the training set.
Specifically, for each positive sample, the real image is used as an input image, and a reconstructed image is generated after the input image passes through a generator network; inputting the reconstructed image into a discriminator network, and outputting a reconstruction score and a reconstruction characteristic; the real image is used as an input image, and a real score and a real feature are output after the real image directly passes through the discriminator network; obtaining a minimized loss function according to the real image, the reconstructed image, the real feature, the reconstructed feature, the real score and the reconstructed score.
The minimized loss comprises a total generator loss and a total discriminator loss; the total loss of the generator comprises residual loss, countermeasure loss and characteristic loss; the total loss of the discriminator comprises real graph confrontation loss and reconstructed graph confrontation loss; and the countermeasure training module alternately trains the corresponding network through back propagation according to the total loss of the generator and the total loss of the discriminator.
And 3, updating the defect detection model by utilizing the generator network and the discriminator network obtained by training to obtain an updated defect detection model.
Specifically, after each round of training is finished, testing is carried out through a verification set, and an AUC value is calculated to serve as a test index of the round of training; and after all training is finished, storing the generator network and the discriminator network corresponding to the highest AUC value, and using the generator network and the discriminator network in the updated defect detection model.
And 4, collecting the image of the panel to be detected.
And 5, preprocessing the panel image to be detected.
Specifically, the pretreatment comprises: extracting an ROI (region of interest) region of the panel image to be detected to obtain an ROI region map; segmenting the ROI regional graph to obtain a plurality of first images; segmenting each first image to obtain a plurality of second images serving as batch processing images; and inputting the batch processing images into the updated defect detection model to obtain the defect detection information corresponding to each second image.
And 6, inputting the image of the panel to be detected into the updated defect detection model, detecting the defect and obtaining defect detection information.
And 7, defect classification is carried out.
Specifically, known defect information is obtained according to the panel image to be detected; and judging whether the defect is a known defect or an abnormal defect by using the overlapping area according to the defect detection information and the known defect information.
The present invention is further described below.
Based on the above object, the present invention uses the recently emerging GAN to perform unsupervised counterstudy only on the positive sample of the industrial display panel, and uses the synthesized output result of the generator and the discriminator of the learned GAN to score each ROI region of the panel and perform abnormal segmentation on possible abnormal regions to obtain abnormal contours. Specifically, the present invention utilizes a generator network of anomally GAN, with a discriminator network as shown in fig. 1 and 2, respectively.
The following description is made for the above-described fig. 1 and 2:
(1) the network shape of the generator is called Unet due to its U-shape. The invention ensures that the Unet structure is dynamically adjusted according to the size of the input image by modifying the script generated by the network, and particularly, the larger the size of the input image is, the more the number of network down-sampling layers is, and the more the corresponding number of up-sampling layers is. Fig. 1 and 2 show the corresponding structure of the network when the input image is 128x 128.
(2) The down-sampling comprises three sub-processing processes, namely a LeakyRelu layer, a convolution layer and a batch normalization layer (BatchNorm) in sequence, wherein the input sequentially passes through the three sub-layers, and then the output is used as the input of the next layer to continue forward propagation.
(3) The up-sampling also has three sub-processing processes, namely a Relu layer, a transposed convolution layer and a batch normalization layer (BatchNorm) in sequence, wherein the input passes through the three sub-layers in sequence, and then the output is used as the input of the next layer for further forward propagation.
(4) Whether input or output, the form is a four-dimensional tensor, taking the input image of the generator as an example: the first dimension of the tensor represents the number of pictures processed in one batch, the illustration is 64, namely the number of pictures processed in one batch is 64; the second dimension represents the number of channels of each picture, and the illustration is 3, which represents that each picture is three channels (such as RGB); the third and fourth represent the height and width dimensions of each picture, illustrated as 128 and 128, representing a 128x128 size for each picture. The rest four-dimensional tensors are analogized in turn.
(5) For the generator, the tensor output by the generator is the reconstruction of the input image, and is called as a reconstructed image; for the discriminator, the output tensor is different from other dimensions, and only has two dimensions, the first dimension still represents the number of batch processing pictures, and the illustration is 64; while the second dimension represents the probability that the input image is normal, the network of discriminators should strive to make the output of normal images approach 1 and the output of abnormal images approach 0.
(6) The generator is essentially an Auto Encoder-Decoder structure network, encoding is carried out by down sampling, decoding is carried out by up sampling, and essentially a generating model; the above-mentioned discriminator is essentially a discriminant model directly after fitting the posterior probability.
The above only illustrates the separate structures of the generator and the network of discriminators, respectively, while fig. 3 shows the overall structure diagram of the two cascaded for antagonism training.
The following is illustrated for fig. 3:
(1) fig. 3 shows a batch of 64 images (input tensors) of 3 channels 128x128 forward propagation process, where Real _ Image first passes through a generator to generate a corresponding reconstructed Image (Fake _ Image), then this reconstructed Image is forward propagated as an input of a discriminator, and finally the discriminator outputs Fake _ Score (last layer in fig. 2) and Fake _ Feature (second last layer in fig. 2) of Fake _ Image.
(2) The Real _ Image not only walks along the forward propagation path to generate the Score and the Feature of the Fake _ Image, but also directly inputs the Score and the Feature to the discriminator to obtain Real _ Score and Real _ Feature. The forward propagation of the entire network system is thus terminated.
(3) The 6 input or output tensors of Real _ Image, Fake _ Image, Real _ Feature, Fake _ Feature, Real _ Score and Fake _ Score are used as the loss functions to be optimized, which are adopted by the patent, are as follows:
total loss of generators for generator needsAs small as possible, the correlation expression is as follows:
wherein,in order to generate the loss of the residual error of the generator,in order for the generator to combat the loss,in order to generate a loss of the characteristics of the device,in order to lose the weight coefficient for the residual error,in order to combat the loss of the weight coefficients,losing weight for featuresThe coefficients of which are such that,score_onesis a tensor with elements all 1.
There is a need for an arbiter to have total loss of the arbiterAs small as possible, the correlation expression is as follows:
wherein,in order for the discriminator real image to combat the loss,for the discriminator to reconstruct the image immunity loss, cross _ entropy represents the cross entropy,score_onesis a tensor in which the elements are all 1,score_zerosis a tensor with elements all 0.
For the above formulas (1) to (7), the following description is given:
(1) g (generator) represents a generator, and d (discriminator) represents a discriminator.
(2) The total loss function of the generator is composed of three parts, namely residual loss, adaptive loss and feature loss, and has three corresponding weight coefficients. Residual loss is the average value of pixel values of a reconstructed image and a real image, and for the GAN only learning a positive sample, the residual loss is minimized so that the model parameters can learn the coding of a normal sample; the penalty is the score of the reconstructed image by the arbiter and the cross entropy of the vector with all elements 1, which means that the reconstructed image produced by the generator is to be perceived by the arbiter as a real picture, i.e. in order for the generator to try to trick the arbiter, the penalty must be minimized; feature loss is the mean square error (mean square error) of the features extracted by the discriminator from the real graph and reconstructed image, which is L2 loss, and also the feature loss must be minimized in order for the reconstructed image produced by the generator to be considered by the discriminator as a real graph.
(3) The total loss function of the discriminator consists of two parts, namely real graph confrontation loss and reconstructed graph confrontation loss. To combat the generator, the arbiter needs to ensure that the real graph is determined to be true and the reconstructed graph is determined to be false, so it is ensured that the cross entropy of the real icon label and all 1 (closer to 1 indicates more true) vectors is minimized, and that the cross entropy of the reconstructed graph label and all 0 (closer to 0 indicates more false) vectors is minimized, which is minimized.
After the total loss of the generator and the discriminator is calculated respectively, the generator and the discriminator need to carry out back propagation according to the loss, wherein the generator and the discriminator are divided into two stages, firstly, the discriminator needs to carry out back propagation by utilizing a gradient descent method from the total loss of the discriminator, and meanwhile, the model parameters of the generator are ensured to be in a frozen state in the propagation process; and after the discriminant completes one batch of training, immediately freezing the model parameters of the discriminant, and reversely propagating the generator by using the total loss function of the generator and a gradient descent method. And finishing all the training rounds.
In the model training process, a test needs to be performed at certain batch intervals, and a test set comprises positive samples and negative samples (defect samples) in the industrial display panel. In any sample, the model has a score for each sample, the score indicates the abnormal score of the sample, the larger the value of the abnormal score, the larger the difference between the normal sample and the sample, and the score expression is calculated as follows:
wherein,andrepresenting the specific gravity of residual loss and feature loss, respectively, in the total anomaly score. Specifically, it is preferable=0.9、=0.1。
As can be seen from the above equation, the larger the residual loss and the characteristic loss of a certain sample, the higher the degree of abnormality, and the larger the residual loss accounts for the characteristic loss.
In the testing process, the trained model is used for scoring the positive samples and the negative samples of the verification set, and the ROC (Receiver Operating Characteristic) curves corresponding to the scores are obtained and the AUC (area under Current) value is obtained, so that the larger the AUC value is, the stronger the capability of distinguishing the normal samples from the abnormal samples is. And finally, selecting the trained model with the highest AUC value as an abnormality detection model in the field industrial panel sample.
For the defect detection GAN of the invention, the abnormal degree of the sample can be accurately scored, and the defect area can be roughly positioned by utilizing the approximate distribution of the pixel values of a residual image (the difference between an original image and a reconstructed image) and matching with morphological operation and an edge detection algorithm.
The whole detection flow will be described below.
For real-time industrial anomaly detection, because the size of a panel is generally very large (6000 x 15000), and meanwhile, when a high-precision camera is used for image shooting by AOI, a non-ROI area is shot together, therefore, if the defect detection GAN network (namely a defect detection model) is used for detecting the anomaly area of the large-area panel, a plurality of preprocessing modules are required to be added; meanwhile, if the position of the abnormal area needs to be positioned more accurately, one score is far from enough, the original image needs to be divided into areas, and each sub-area needs to be scored abnormally; if the complete contour of the abnormal region needs to be determined more accurately, not only the abnormal score but also the residual map and the morphological algorithm need to be used for further processing.
The invention provides a set of high-efficiency industrial abnormal defect detection flow scheme, as shown in figure 4. The processing flow of the anomaly detection module is shown in fig. 5.
The following is illustrated for this flow:
(1) the resolution of the raw image of a high-precision camera is very high and is typically cut into several large sub-blocks which are then processed separately.
(2) The ROI extraction algorithm is generally designed according to different prior information of different sites, is different for each site, and needs to be specifically designed for each site.
(3) The ROI is further divided into medium size maps with smaller resolution, typically 1024x1024, 512x512 or other medium size, 512x512 is taken as an example in fig. 4.
(4) The 512x512 graph needs to be cut again before anomaly detection, so as to more accurately locate the anomaly region. If the image size of the training defect inspection GAN is 128 × 128, the width and height are divided into 4 parts each with a size of 128, and 16 parts in total, and forward propagation is performed simultaneously as one batch process instead of sequentially.
(5) After the anomaly detection is carried out forward, not only the score of each of the 16 sub-regions is obtained, but also a residual image of the whole original image (namely 16 images are spliced into the original image according to the original positions) is obtained, the anomaly region can be more accurately positioned after the morphological and contour extraction is carried out, and the anomaly detection module outputs three kinds of information: 1. whether each patch is abnormal, the pool value; 2. a list of outlines list; 3. the corresponding bounding box list.
(6) The comprehensive judgment module is positioned at the upper layer of the algorithm, the information of the known defects obtained by the detection module and the three kinds of information output by the abnormity detection module are required to be utilized, the overlapped area is utilized to comprehensively judge whether a certain sub-region is abnormal or the known defects, and different marks are used for representing the sub-region. Where the detection module detects known defects using conventional defect detection methods, such as supervised learning defect detection methods.
The defect detection model and the whole abnormal defect detection method have been tested on an actual field sample set of the display panel and have good effects, and fig. 6 shows an abnormal defect detection result sample diagram of the LCD display panel appearance inspection provided by a certain panel manufacturer.
The description of fig. 6 is as follows:
(1) the original image has four defects, three of which are known defects and the other is abnormal defect (small point in the image)
(2) The residual image between the original image and the reconstructed image obtained after the original image is subjected to forward propagation by the anomaly detection module is also three channels, and then the residual image on the left side of the image 6 is obtained after graying and threshold transformation, so that the pixel value obtained after abnormal pixel points (including all defective pixel points of known defects and abnormal defects) in the original image are subjected to defect detection GAN is obviously different from the original abnormal position, and the difference of the normal position is not large, so that the difference is not more than the threshold value.
(3) The right diagram is a bounding box (bounding box) corresponding to each continuous abnormal region obtained after a simple closing operation, a contour extraction algorithm and a contour filtering module, and it can be seen that there are 4 bounding boxes in the list, where the bounding boxes 1 to 3 frame out scratches (scratch) on the screen, that is, the first three corresponding defects are known defects on site, and the fourth bounding box is a defect that is not present in the sample set, that is, an abnormal defect. The anomaly detection algorithm is a single type learning method based on positive samples, and only knows that the 4 boxes are all defect boxes, but cannot know the specific defects. Assuming that the same image passes through the detection module, the detection network training data corresponding to the detection module has scratches (scratch) therein, but does not have data with the same characteristics as the defects framed by the bounding box 4, the detection module obtains 3 detection frames according to the known data set training, and determines which of the 4 bounding boxes of the right image are known defects and which are unknown defects by using the overlapping areas of the 4 bounding boxes of the right image and the 3 detection frames obtained by the detection module. Alternatively, an IOU (Intersection over Union) may be used as an index for measuring the overlapping area. For example, the detection module trains according to a known data set to obtain the positions of the 3 detection frames and the IOU (Intersection over Union) of the first 3 detection frames for abnormal detection, but the IOU of the fourth detection frame for abnormal detection and the IOU of the 3 detection frames for abnormal detection are not large (although the boundary frame 4 is inside the boundary frame 2, the IOU is not large), so that the integration module can determine that the boundary frames 1-3 are known defects and the boundary frame 4 corresponds to an abnormal defect. For example, the first abnormal detection box (i.e., the boundary box 1) and the 3 detection boxes IOU are (0.9, 0.01, 0.01), respectively, and 0.9 is greater than the preset threshold, and thus it is determined that the defect is known, and the second and third abnormal detection boxes are similar, and the fourth abnormal detection box is reached and the IOU is (0.02, 0.02, 0.02), neither is greater than the threshold, and thus it is determined that the defect in the fourth abnormal detection box is an abnormal defect rather than a known defect. For a plurality of practical field applications, it is shown that the defect detection GAN based on the present invention has a better discrimination between the defect sample and the normal sample, and because the input tensor is a batch sample, even if a graph is divided into a plurality of sub-regions (for example, 512 × 512 is divided into 16 sub-regions), the parallel forward propagation can be performed by using the parallelism of the graphics card, and the real-time detection effect can be actually achieved, and a large number of test results of practical field are shown in table 1:
TABLE 1 statistics of sub-region area, missed-detection rate, and average single-picture time consumption
patch size | miss rate | mistake rate | avg time / image |
64x64 | 0.40% | 2.75% | 138.5ms |
128x128 | 6.02% | 3.52% | 70.3ms |
256x256 | 13.03% | 2.61% | 30.1ms |
The single graph here is fixed with a 512x512 size, and is measured with this size as a reference. It is understood that the original image can be divided into 8x8 parts by 64x64, 4x4 parts by 128x128, and 2x2 parts by 256x 256.
Wherein miss rate indicates that this sub-area is actually defective or part of a defect but is judged to be normal; mistake rate means that this sub-region is originally a positive sample, but is misjudged as defective; the average single graph time consumption shows that the time consumption is smaller when the patch size is larger, but the detection precision is correspondingly reduced at the same time, particularly the omission ratio.
According to the real-time requirement of the industrial field, the performance of the invention can meet the real-time requirement of most industrial fields after the bottom layer tensorrT + FP16 is further accelerated and optimized.
The AOI defect detection method provided by the embodiment of the invention at least comprises the following technical effects:
1. the invention can obtain the defect detection information based on the positive sample unsupervised training GAN network, and effectively distinguish the normal sample from the defect sample.
2. The method is based on the positive sample unsupervised training GAN network, can effectively distinguish the defects into known defects or abnormal defects, effectively solves the detection problem of the unknown defects (abnormal defects), and meets the detection requirement of industrial real-time property.
3. The invention provides a general defect detection process, an abnormal defect detection process and a GAN detection training process, which are suitable for various different fields. Meanwhile, the method is based on only positive samples, and the number of the positive samples is very large, so that the cost of data acquisition is greatly reduced.
Finally, it should be noted that the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting, and although the present invention has been described in detail with reference to examples, it should be understood by those skilled in the art that modifications or equivalent substitutions may be made on the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, which should be covered by the claims of the present invention.
Claims (10)
1. An AOI defect detection method is characterized by comprising the following steps:
establishing a defect detection model, wherein the defect detection model comprises a generator network and a discriminator network, and is used for obtaining defect related information, and the defect related information comprises an abnormal score of an input sample and/or specific position information of an abnormal area;
establishing a training set by using positive samples, and training the generator network and the discriminator network by using the training set;
updating the defect detection model by using the generator network and the discriminator network obtained by training to obtain an updated defect detection model;
collecting an image of a panel to be detected;
and inputting the image of the panel to be detected into the updated defect detection model, detecting the defects, and obtaining the related information of the defects of the panel to be detected.
2. The AOI defect detection method of claim 1, wherein the specific location information of the abnormal region comprises an abnormal region outline and/or an abnormal region bounding box.
3. The AOI defect detection method according to claim 1, wherein the generator network adopts a Unet structure, and an input image is subjected to N-layer (N > = 1) down-sampling and N-layer up-sampling to obtain a reconstructed image;
the down-sampling comprises three sub-processing processes, namely a LeakyRelu layer, a convolution layer and a batch normalization layer in sequence; the up-sampling comprises three sub-processing processes, namely a Relu layer, a transposed convolution layer and a batch normalization layer.
4. The AOI defect detection method according to claim 1, wherein the input image of the discriminator network is subjected to N (N > = 1) layers of downsampling, feature extraction, and convolution processing to obtain an output score;
the down-sampling and the feature extraction both comprise three sub-processing processes, namely a LeakyRelu layer, a convolution layer and a batch normalization layer in sequence; the convolution processing comprises two sub-processing processes, namely a convolution layer and a sigmoid layer.
5. The AOI defect detection method according to claim 2, wherein the defect detection model comprises an abnormal GAN forward module and a feature information extraction module connected in sequence;
the abnormal GAN forward module comprises a generator network and a discriminator network which are connected in sequence and is used for obtaining the abnormal score and the residual map of the input sample;
the characteristic information extraction module comprises a residual error map opening and closing operation module, an abnormal region outline extraction module and an outline filtering module which are sequentially connected and is used for obtaining an abnormal region outline and an abnormal region boundary frame according to the residual error map.
6. The AOI defect detection method of claim 1, wherein the training the generator network and the discriminator network with the training set is:
for each positive sample, the real image is used as an input image, and a reconstructed image is generated after the input image passes through a generator network; inputting the reconstructed image into a discriminator network, and outputting a reconstruction score and a reconstruction characteristic; the real image is used as an input image, and a real score and a real feature are output after the real image directly passes through the discriminator network;
obtaining a minimized generator loss function and a discriminator loss function from the real image, the reconstructed image, the real feature, the reconstructed feature, the real score, and the reconstructed score.
7. The AOI defect detection method of claim 6, wherein the generator loss function comprises residual loss, countermeasure loss, feature loss; the discriminator loss function comprises real graph confrontation loss and reconstructed graph confrontation loss;
and training the corresponding network by back propagation alternately according to the generator loss function and the discriminator loss function by utilizing the training set.
8. The AOI defect detection method of claim 1, wherein after each round of training is completed, the test is performed through a validation set, and an AUC value is calculated as a test index of the round of training; and after all training is finished, storing the generator network and the discriminator network corresponding to the highest AUC value, and using the generator network and the discriminator network in the updated defect detection model.
9. The AOI defect inspection method of claim 1, further comprising, before inputting the panel image to be inspected into the updated defect inspection model: preprocessing the panel image to be detected;
the pretreatment comprises the following steps: extracting an ROI (region of interest) region of the panel image to be detected to obtain an ROI region map; segmenting the ROI regional graph to obtain a plurality of first images; segmenting each first image to obtain a plurality of second images serving as batch processing images;
and inputting the batch processing images into the updated defect detection model to obtain defect related information corresponding to each second image.
10. The AOI defect inspection method of any one of claims 1-9, further comprising performing defect classification, specifically:
obtaining known defect information according to the panel image to be detected; and judging whether the defect is a known defect or an abnormal defect by using the overlapping area according to the defect related information and the known defect information.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010141504.0A CN111223093A (en) | 2020-03-04 | 2020-03-04 | AOI defect detection method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010141504.0A CN111223093A (en) | 2020-03-04 | 2020-03-04 | AOI defect detection method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111223093A true CN111223093A (en) | 2020-06-02 |
Family
ID=70829936
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010141504.0A Pending CN111223093A (en) | 2020-03-04 | 2020-03-04 | AOI defect detection method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111223093A (en) |
Cited By (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111693534A (en) * | 2020-06-12 | 2020-09-22 | 北京百度网讯科技有限公司 | Surface defect detection method, model training method, device, equipment and medium |
CN111724372A (en) * | 2020-06-19 | 2020-09-29 | 深圳新视智科技术有限公司 | Method, terminal and storage medium for detecting cloth defects based on antagonistic neural network |
CN112116557A (en) * | 2020-08-12 | 2020-12-22 | 西安交通大学 | Radiographic image weld area defect detection method, storage medium and equipment |
CN112288711A (en) * | 2020-10-28 | 2021-01-29 | 浙江华云清洁能源有限公司 | Unmanned aerial vehicle inspection image defect image identification method, device, equipment and medium |
CN112508946A (en) * | 2020-12-29 | 2021-03-16 | 苏州光格科技股份有限公司 | Cable tunnel abnormity detection method based on antagonistic neural network |
CN112669274A (en) * | 2020-12-23 | 2021-04-16 | 山东大学 | Multi-task detection method for pixel-level segmentation of surface abnormal region |
CN113052203A (en) * | 2021-02-09 | 2021-06-29 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Anomaly detection method and device for multiple types of data |
CN113379689A (en) * | 2021-05-31 | 2021-09-10 | 东南大学 | Defect detection algorithm based on deep learning |
CN113688889A (en) * | 2021-08-13 | 2021-11-23 | 上海商汤智能科技有限公司 | Abnormality detection method, abnormality detection device, electronic apparatus, and computer-readable storage medium |
CN113822888A (en) * | 2021-11-23 | 2021-12-21 | 中导光电设备股份有限公司 | Automatic division method and system based on liquid crystal screen detection area |
CN113837209A (en) * | 2020-06-23 | 2021-12-24 | 乐达创意科技股份有限公司 | Method and system for improved machine learning using data for training |
CN113838015A (en) * | 2021-09-15 | 2021-12-24 | 上海电器科学研究所(集团)有限公司 | Electric appliance product appearance defect detection method based on network cooperation |
CN113838024A (en) * | 2021-09-22 | 2021-12-24 | 哈尔滨工业大学 | OLED panel defect prediction method |
WO2022067876A1 (en) * | 2020-09-29 | 2022-04-07 | Hong Kong Applied Science And Technology Research Institute Co. Ltd. | Method and system of image based anomaly localization for vehicles through generative contextualized adversarial network |
CN115661160A (en) * | 2022-12-29 | 2023-01-31 | 成都数之联科技股份有限公司 | Panel defect detection method, system, device and medium |
CN115880290A (en) * | 2023-02-22 | 2023-03-31 | 季华实验室 | OLED (organic light emitting diode) wet film defect detection method based on lightweight semantic segmentation network |
CN117094986A (en) * | 2023-10-13 | 2023-11-21 | 中山大学深圳研究院 | Self-adaptive defect detection method based on small sample and terminal equipment |
CN117291922A (en) * | 2023-11-27 | 2023-12-26 | 浙江日井泵业股份有限公司 | Visual detection method for defects of stainless steel multistage pump impeller |
CN117315354A (en) * | 2023-09-27 | 2023-12-29 | 南京航空航天大学 | Insulator anomaly detection method based on multi-discriminant composite coding GAN network |
CN117437202A (en) * | 2023-11-01 | 2024-01-23 | 深圳市宇创显示科技有限公司 | Image-based polaroid defect detection method, device, equipment and storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110097543A (en) * | 2019-04-25 | 2019-08-06 | 东北大学 | Surfaces of Hot Rolled Strip defect inspection method based on production confrontation network |
CN110322433A (en) * | 2019-05-27 | 2019-10-11 | 苏州佳赛特智能科技有限公司 | A kind of data set amplification method towards the detection of appearance defective vision |
CN110598851A (en) * | 2019-08-29 | 2019-12-20 | 北京航空航天大学合肥创新研究院 | Time series data abnormity detection method fusing LSTM and GAN |
US10565477B2 (en) * | 2016-11-23 | 2020-02-18 | General Electric Company | Deep learning medical systems and methods for image reconstruction and quality evaluation |
-
2020
- 2020-03-04 CN CN202010141504.0A patent/CN111223093A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10565477B2 (en) * | 2016-11-23 | 2020-02-18 | General Electric Company | Deep learning medical systems and methods for image reconstruction and quality evaluation |
CN110097543A (en) * | 2019-04-25 | 2019-08-06 | 东北大学 | Surfaces of Hot Rolled Strip defect inspection method based on production confrontation network |
CN110322433A (en) * | 2019-05-27 | 2019-10-11 | 苏州佳赛特智能科技有限公司 | A kind of data set amplification method towards the detection of appearance defective vision |
CN110598851A (en) * | 2019-08-29 | 2019-12-20 | 北京航空航天大学合肥创新研究院 | Time series data abnormity detection method fusing LSTM and GAN |
Non-Patent Citations (2)
Title |
---|
Y. T. K. LAI ETC: "Industrial Anomaly Detection and One-class Classification using Generative Adversarial Networks", 《PROCEEDINGS OF THE 2018 IEEE/ASME INTERNATIONAL CONFERENCE ON ADVANCED INTELLIGENT MECHATRONICS (AIM)》 * |
李卓蓉: "生成式对抗网络研究及其应用", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (36)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111693534A (en) * | 2020-06-12 | 2020-09-22 | 北京百度网讯科技有限公司 | Surface defect detection method, model training method, device, equipment and medium |
CN111693534B (en) * | 2020-06-12 | 2023-09-15 | 北京百度网讯科技有限公司 | Surface defect detection method, model training method, device, equipment and medium |
US11763552B2 (en) | 2020-06-12 | 2023-09-19 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method for detecting surface defect, method for training model, apparatus, device, and media |
CN111724372A (en) * | 2020-06-19 | 2020-09-29 | 深圳新视智科技术有限公司 | Method, terminal and storage medium for detecting cloth defects based on antagonistic neural network |
WO2021253632A1 (en) * | 2020-06-19 | 2021-12-23 | 深圳新视智科技术有限公司 | Cloth defect detection method based on adversarial neural network, and terminal and storage medium |
CN113837209A (en) * | 2020-06-23 | 2021-12-24 | 乐达创意科技股份有限公司 | Method and system for improved machine learning using data for training |
CN112116557A (en) * | 2020-08-12 | 2020-12-22 | 西安交通大学 | Radiographic image weld area defect detection method, storage medium and equipment |
CN112116557B (en) * | 2020-08-12 | 2022-02-22 | 西安交通大学 | Radiographic image weld area defect detection method, storage medium and equipment |
US11328402B2 (en) | 2020-09-29 | 2022-05-10 | Hong Kong Applied Science and Technology Research Institute Company Limited | Method and system of image based anomaly localization for vehicles through generative contextualized adversarial network |
WO2022067876A1 (en) * | 2020-09-29 | 2022-04-07 | Hong Kong Applied Science And Technology Research Institute Co. Ltd. | Method and system of image based anomaly localization for vehicles through generative contextualized adversarial network |
CN112288711A (en) * | 2020-10-28 | 2021-01-29 | 浙江华云清洁能源有限公司 | Unmanned aerial vehicle inspection image defect image identification method, device, equipment and medium |
CN112669274A (en) * | 2020-12-23 | 2021-04-16 | 山东大学 | Multi-task detection method for pixel-level segmentation of surface abnormal region |
CN112669274B (en) * | 2020-12-23 | 2022-06-17 | 山东大学 | Multi-task detection method for pixel-level segmentation of surface abnormal region |
CN112508946A (en) * | 2020-12-29 | 2021-03-16 | 苏州光格科技股份有限公司 | Cable tunnel abnormity detection method based on antagonistic neural network |
CN112508946B (en) * | 2020-12-29 | 2024-04-26 | 苏州光格科技股份有限公司 | Cable tunnel anomaly detection method based on antagonistic neural network |
CN113052203A (en) * | 2021-02-09 | 2021-06-29 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Anomaly detection method and device for multiple types of data |
CN113052203B (en) * | 2021-02-09 | 2022-01-18 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Anomaly detection method and device for multiple types of data |
US11423260B1 (en) | 2021-02-09 | 2022-08-23 | Harbin Institute Of Technology (Shenzhen) (Shenzhen Institute Of Science And Technology Innovation, Harbin Institute Of Technology) | Anomaly detection method and apparatus for multi-type data |
CN113379689B (en) * | 2021-05-31 | 2024-05-10 | 东南大学 | Defect detection algorithm based on deep learning |
CN113379689A (en) * | 2021-05-31 | 2021-09-10 | 东南大学 | Defect detection algorithm based on deep learning |
CN113688889A (en) * | 2021-08-13 | 2021-11-23 | 上海商汤智能科技有限公司 | Abnormality detection method, abnormality detection device, electronic apparatus, and computer-readable storage medium |
CN113838015A (en) * | 2021-09-15 | 2021-12-24 | 上海电器科学研究所(集团)有限公司 | Electric appliance product appearance defect detection method based on network cooperation |
CN113838015B (en) * | 2021-09-15 | 2023-09-22 | 上海电器科学研究所(集团)有限公司 | Electrical product appearance defect detection method based on network cooperation |
CN113838024A (en) * | 2021-09-22 | 2021-12-24 | 哈尔滨工业大学 | OLED panel defect prediction method |
CN113822888A (en) * | 2021-11-23 | 2021-12-21 | 中导光电设备股份有限公司 | Automatic division method and system based on liquid crystal screen detection area |
CN115661160A (en) * | 2022-12-29 | 2023-01-31 | 成都数之联科技股份有限公司 | Panel defect detection method, system, device and medium |
CN115880290A (en) * | 2023-02-22 | 2023-03-31 | 季华实验室 | OLED (organic light emitting diode) wet film defect detection method based on lightweight semantic segmentation network |
WO2024174349A1 (en) * | 2023-02-22 | 2024-08-29 | 季华实验室 | Oled wet film defect detection method based on lightweight semantic segmentation network |
CN117315354A (en) * | 2023-09-27 | 2023-12-29 | 南京航空航天大学 | Insulator anomaly detection method based on multi-discriminant composite coding GAN network |
CN117315354B (en) * | 2023-09-27 | 2024-04-02 | 南京航空航天大学 | Insulator anomaly detection method based on multi-discriminant composite coding GAN network |
CN117094986A (en) * | 2023-10-13 | 2023-11-21 | 中山大学深圳研究院 | Self-adaptive defect detection method based on small sample and terminal equipment |
CN117094986B (en) * | 2023-10-13 | 2024-04-05 | 中山大学深圳研究院 | Self-adaptive defect detection method based on small sample and terminal equipment |
CN117437202A (en) * | 2023-11-01 | 2024-01-23 | 深圳市宇创显示科技有限公司 | Image-based polaroid defect detection method, device, equipment and storage medium |
CN117437202B (en) * | 2023-11-01 | 2024-07-16 | 深圳市宇创显示科技有限公司 | Image-based polaroid defect detection method, device, equipment and storage medium |
CN117291922A (en) * | 2023-11-27 | 2023-12-26 | 浙江日井泵业股份有限公司 | Visual detection method for defects of stainless steel multistage pump impeller |
CN117291922B (en) * | 2023-11-27 | 2024-01-30 | 浙江日井泵业股份有限公司 | Visual detection method for defects of stainless steel multistage pump impeller |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111223093A (en) | AOI defect detection method | |
CN111223088B (en) | Casting surface defect identification method based on deep convolutional neural network | |
CN108562589B (en) | Method for detecting surface defects of magnetic circuit material | |
CN111383209B (en) | Unsupervised flaw detection method based on full convolution self-encoder network | |
CN111652085B (en) | Object identification method based on combination of 2D and 3D features | |
CN109191421B (en) | Visual detection method for pits on circumferential surface of cylindrical lithium battery | |
CN107966454A (en) | A kind of end plug defect detecting device and detection method based on FPGA | |
CN111402226A (en) | Surface defect detection method based on cascade convolution neural network | |
CN111047655B (en) | High-definition camera cloth defect detection method based on convolutional neural network | |
CN113920107A (en) | Insulator damage detection method based on improved yolov5 algorithm | |
CN113643268B (en) | Industrial product defect quality inspection method and device based on deep learning and storage medium | |
CN111382785A (en) | GAN network model and method for realizing automatic cleaning and auxiliary marking of sample | |
CN112233067A (en) | Hot rolled steel coil end face quality detection method and system | |
CN112053317A (en) | Workpiece surface defect detection method based on cascade neural network | |
CN110473201A (en) | A kind of automatic testing method and device of disc surface defect | |
CN114881987B (en) | Hot-pressing light guide plate defect visual detection method based on improvement YOLOv5 | |
CN110009622B (en) | Display panel appearance defect detection network and defect detection method thereof | |
CN116990323B (en) | High-precision printing plate visual detection system | |
CN115830004A (en) | Surface defect detection method, device, computer equipment and storage medium | |
CN113962929A (en) | Photovoltaic cell assembly defect detection method and system and photovoltaic cell assembly production line | |
CN116612106A (en) | Method for detecting surface defects of optical element based on YOLOX algorithm | |
CN116205876A (en) | Unsupervised notebook appearance defect detection method based on multi-scale standardized flow | |
CN115205209A (en) | Monochrome cloth flaw detection method based on weak supervised learning | |
CN117252815A (en) | Industrial part defect detection method, system, equipment and storage medium based on 2D-3D multi-mode image | |
CN115866502A (en) | Microphone part surface defect online detection process |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200602 |
|
RJ01 | Rejection of invention patent application after publication |