Mini-LED Defect Detection Method and System Based on Wide Learning

Abstract

The present invention provides a Mini-LED defect detection method and system based on width learning, which relates to the technical field of display product defect detection. The method includes: obtaining an initial image of a Mini-LED product and performing preprocessing, inputting the preprocessed initial image of the Mini-LED product into a width learning defect detection neural network model with incremental learning for training, and using the optimized width learning defect detection neural network model with incremental learning to perform defect detection on the image of the Mini-LED product to be detected; The present invention uses a width learning neural network with incremental learning for defect detection, which has a flat structure, simple calculation, and fast training speed. The incremental learning improves the accuracy of defect detection by adding enhancement nodes; In addition, when the model is trained, there is no large amount of calculation, which can significantly improve the efficiency and accuracy of Mini-LED product defect detection and classification.

Description

Technical Field

[0001] The present invention relates to the technical field of display product defect detection, and more specifically, to a Mini-LED defect detection method and system based on width learning. Background Art

[0002] Different from the size of conventional LED products, the size of Mini-LEDs is generally between 50-200 μm, and it mainly consists of a Mini-LED pixel array and a driving circuit, and is composed of units with a pixel center pitch of 0.3-1.5 mm. In recent years, Mini-LED display products have been widely used in ultra-large screen high-definition displays, such as commercial fields such as monitoring and command, high-definition studio, high-end cinema, medical diagnosis, advertising display, conference and exhibition, office display, virtual reality, etc.

[0003] With the development of new industries and new technologies, people's requirements for product quality, especially display screens, are getting higher and higher. Quality inspection of display screens during the production process is an essential process. The optical resolution of the conventional LED defect detection method fails to meet the detection requirements of Mini-LEDs. Currently, the commonly used Mini-LED defect detection methods mainly include three categories: artificial vision detection method, electrical detection method, and optical detection method. The artificial vision detection method is a relatively primitive surface defect detection method, that is, a microscope and a special light source are used to detect product defects by the naked eye. The disadvantages are slow detection speed, low efficiency, and low detection accuracy, and it is easy to miss detections, unable to meet the production capacity requirements; the commonly used electrical detection methods include full-screen lighting method, charge reading method, electron beam scanning pixel electrode method, probe scanning method, voltage image method, and admittance circuit detection method, etc. However, because it can only detect defects caused by electrical factors, it can only be used for detection after the panel manufacturing is completed, and it cannot detect functional defects during its manufacturing process; the optical detection method can be divided into image processing method and image recognition method. After feature extraction and processing of the acquired image, it is input into a classifier for panel image classification. Because of its advantages such as high precision, high accuracy, no labor cost, and automated operation, it has gradually become the mainstream research direction.

[0004] At the present stage, the research on the optical detection method mainly includes machine vision methods and deep learning methods. The traditional machine vision-based surface defect detection methods mainly include image feature detection and template matching. It is difficult for traditional machine vision methods to extract defect features with high effectiveness, and the data volume is huge. The real-time performance of extracting defect information from a large amount of data is not high, and it is still difficult to meet the current requirements in practical applications.

[0005] The surface defect detection algorithm based on deep learning mainly identifies defects by extracting defect features from pixels. The main methods include Support Vector Machines (SVM) and Convolutional Neural Networks (CNN). Although the deep learning method ensures accuracy, the extraction of its data features is complex and requires a large number of parameter learning for data, resulting in a high time cost. The training process requires a large amount of labeled data to train its model to obtain an effective model, which has little applicability to the actual industrial scenario.

[0006] Although the defect detection method for Mini-LED products based on deep learning can achieve a high accuracy rate, as the number of network layers deepens, this method is often deeply involved in the process of updating weights layer by layer. It not only requires a large amount of training time but also is accompanied by problems such as gradient disappearance, gradient explosion, or local optimal solutions. At the same time, its data demand is large, and it is difficult to obtain a large amount of defect annotation data in actual industrial applications.

[0007] The current existing technology discloses a defect detection method and related equipment for Mini-LED products, including: collecting an image of a Mini-LED product and using the image of the Mini-LED product as a detection image; selecting a template feature database corresponding to the Mini-LED product, inputting the detection image into the template feature database corresponding to the Mini-LED product, and obtaining a reference image using a function library; based on the detection image and the reference image, obtaining a residual image using the ET algorithm; marking the pixel points with gray values greater than the threshold in the residual image as defect points, and marking the pixel points with gray values not greater than the threshold as non-defect points. The method in the existing technology detects the appearance defects by obtaining the real image of the actual Mini-LED product, cannot extract relatively complete features, and the detection effect is not good. In addition, the method in the existing technology only uses a fixed and single defect detection program to perform defect detection, cannot perform learning and iterative update, and the robustness and accuracy of the detection will also be greatly reduced. Summary of the Invention

[0008] To overcome the defects of the low robustness and accuracy of the above existing technology in the appearance defect detection of Mini-LED, the present invention provides a Mini-LED defect detection method and system based on width learning, which can improve the detection efficiency and accuracy of the appearance defect detection of Mini-LED.

[0009] To solve the above technical problems, the technical solution of the present invention is as follows:

[0010] A Mini-LED defect detection method based on width learning, comprising the following steps:

[0011] S1: Obtain the initial image of the Mini-LED product and perform preprocessing;

[0012] S2: Establish a width learning defect detection neural network model with incremental learning;

[0013] S3: Input the preprocessed initial image of the Mini-LED product into the width learning defect detection neural network model with incremental learning for training to obtain an optimized width learning defect detection neural network model with incremental learning;

[0014] S4: Obtain the image of the Mini-LED product to be detected, and use the optimized width learning defect detection neural network model with incremental learning to perform defect detection on the image of the Mini-LED product to be detected.

[0015] Preferably, in the step S1, the specific method for obtaining the initial image of the Mini-LED product and performing preprocessing is as follows:

[0016] The types of the initial images of the Mini-LED products include the initial images of normal products, the initial images of products with foreign objects, and the initial images of products with missing lamp beads;

[0017] The specific method of the preprocessing is as follows:

[0018] Divide all the initial images of the Mini-LED products into a training set and a test set, and perform grayscale operation and flattening operation on all the initial images of the Mini-LED products in sequence; fill the pixel values of all pixel points into a table, and each column of pixel values corresponds to an initial image of the Mini-LED product; mark different labels according to the different types of the initial images of the Mini-LED products, fill the labels into the first row of the corresponding table, and perform a random shuffling operation on the pixel information of each column of the table to complete the preprocessing of the initial images of the Mini-LED products.

[0019] Preferably, the size of the initial image of the Mini-LED product is 200*215.

[0020] Preferably, marking different labels according to the different types of the initial images of the Mini-LED products is specifically as follows:

[0021] Mark the initial image of the normal product as label 0, the initial image of the product with foreign objects as label 1, and the initial image of the product with missing lamp beads as label 2.

[0022] Preferably, in the step S2, the width learning defect detection neural network model with incremental learning includes an input layer, a feature layer, an enhancement layer, and an output layer;

[0023] The output end of the input layer is connected to the input end of the feature layer, and the output end of the feature layer is connected to the input end of the enhancement layer;

[0024] The output end of the feature layer is also spliced with the output end of the enhancement layer and jointly connected to the input end of the output layer.

[0025] Preferably, in step S3, the preprocessed initial image of the Mini-LED product is input into the width learning defect detection neural network model with incremental learning for training to obtain an optimized width learning defect detection neural network model with incremental learning. The specific method is as follows:

[0026] S3.1: Use the preprocessed initial image of the Mini-LED product as the input matrix X and input it into the input layer of the width learning defect detection neural network model with incremental learning for calculation to obtain the merged feature node matrix Z n =[Z1, Z2, …, Z n ∈R N×np ;

[0027] S3.2: Calculate and obtain the merged enhancement node matrix H n =[Z1, Z2, …, Z n ∈R N×np according to the merged feature node matrix Z m =[H1, H2, …, H m ;

[0028] S3.3: Merge the merged feature node matrix Z n =[Z1, Z2, …, Z n ∈R N×np with the merged enhancement node matrix H m =[H1, H2, …, H m again to obtain the first merged matrix, denoted as matrix A;

[0029] Matrix A satisfies A = [Z n |H m ;

[0030] S3.4: Add new enhancement nodes and merge them with matrix A to obtain the second merged matrix, denoted as matrix A m+1 ;

[0031] Matrix A m+1 satisfies A m+1 =[A|ξ(Z n W hm+1 +β hm+1 )];

[0032] S3.5: Calculate the pseudo-inverse matrix of matrix A m+1 (Am+1 ) + , according to the matrix A m+1 The pseudo-inverse matrix (A m+1 ) + Get the matrix A m +1 The link weight W between the output matrix Y m+1 ;

[0033] S3.6: Judgment Matrix A m+1 The link weight W between the output matrix Y m+1 Whether the preset error condition is met, if so, the matrix A at this time is m+1 The link weight W between the output matrix Y m+1 Save and obtain the optimized width learning defect detection neural network model with incremental learning; otherwise, repeat steps S3.1 to S3.5.

[0034] Preferably, in step S3.1, the specific method for obtaining the merged feature node matrix is:

[0035] Input the preset input matrix X into the input layer, and after n feature mappings, obtain n sets of feature node matrices Z i , the i-th group feature node matrix Z i Specifically:

[0036] Z i =Φ(XW ei +β ei )∈R N×p ,i=1,2,…,n

[0037] Where X is the input matrix, satisfying X∈R N×M , N represents the number of samples in the input data set, M represents the number of features of the input sample vector; W ei is the first weight matrix with Gaussian distribution, satisfying W ei ∈R M×p β ei is the first bias matrix, satisfying β ei ∈R N×p ; p is the first parameter determined by the grid search method; the function Φ(·) represents a linear feature map;

[0038] The n groups of feature node matrices Z i Merge to obtain the merged feature node matrix, specifically:

[0039] Z n =[Z1,Z2,…,Z n ]∈R N×np

[0040] Among them, Zn Represents the merged feature node matrix.

[0041] Preferably, in the step S3.2, the specific method for obtaining the merged enhanced node matrix is as follows:

[0042] The merged feature node matrix Z n After undergoing m feature mappings, m groups of enhanced node matrices H m are obtained. The m-th group of enhanced node matrix H m is specifically:

[0043] H m = ξ(Z n W hj + β hj ) ∈ R N×q , j = 1, 2, …, m

[0044] where W hj is the second weight matrix with a Gaussian distribution, satisfying W hj ∈ R np×q ; β hj is the second bias matrix, satisfying β hj ∈ R N×q ; q is the second parameter determined by the grid search method; the function ξ(·) represents a non-linear feature mapping;

[0045] The m groups of enhanced node matrices H m are merged to obtain the merged enhanced node matrix, specifically:

[0046] H m = [H1, H2, …, H m ∈ R N×mq

[0047] where H m represents the merged enhanced node matrix.

[0048] Preferably, in the step S3.5, the specific method for obtaining the link weight W m+1 between the matrix A m+1 and the output matrix Y is as follows:

[0049] The feature node matrix Z i of the feature layer and the enhanced node matrix H m of the enhanced layer are merged to obtain the first merged matrix, denoted as matrix A, specifically:

[0050] A = [Z n |H m ∈ R N×L , = np + mq

[0051] The pseudo-inverse matrix A of matrix A+ is:

[0052]

[0053] where λ is the regularization parameter and I is the identity matrix;

[0054] Obtain the link weight W between matrix A and matrix Y according to the pseudo-inverse matrix A + , specifically: m

[0055] W m =(A T A + λI) -1 A T Y

[0056] Obtain a new enhanced node matrix, merge the new enhanced node matrix with matrix A to obtain a second merged matrix, denoted as matrix A m+1 , specifically:

[0057] A m+1 =[A|ξ(Z n W hm+1 + hm+1 )]

[0058] where W hm+1 is the third weight matrix with a Gaussian distribution, satisfying W hm+1 ∈ np×k ; β hm+1 is the third bias matrix, satisfying β hm+1 ∈ k , and k is the third parameter determined by the grid search method;

[0059] The pseudo-inverse matrix of matrix A m+1 ( m+1 ) + is:

[0060]

[0061] where D = + ξ(Z n W hm+1 + hm+1 ); C = ξ(Z n W hm+1 + hm+1 ) - D;

[0062] Obtain the link weight W between matrix A m+1 and the output matrix Y according to the pseudo-inverse matrix of matrix A m+1 ( + ), specifically: m+1 m+1 ​​​

[0063]

[0064] Among them, W m+1 is the link weight between matrix A m+1 and output matrix Y.

[0065] The present invention also provides a Mini-LED defect detection system based on width learning, which applies the above-mentioned Mini-LED defect detection method based on width learning, and includes:

[0066] A preprocessing unit: used to obtain the initial image of the Mini-LED product and perform preprocessing;

[0067] A model construction unit: used to establish a width learning defect detection neural network model with incremental learning;

[0068] A model training unit: used to input the preprocessed initial image of the Mini-LED product into the width learning defect detection neural network model with incremental learning for training, and obtain an optimized width learning defect detection neural network model with incremental learning;

[0069] A defect prediction unit: used to obtain the image of the Mini-LED product to be detected, and perform defect detection on the image of the Mini-LED product to be detected by using the optimized width learning defect detection neural network model with incremental learning.

[0070] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0071] The present invention provides a Mini-LED defect detection method and system based on width learning. The method includes: obtaining the initial image of the Mini-LED product and performing preprocessing; establishing a width learning defect detection neural network model with incremental learning; inputting the preprocessed initial image of the Mini-LED product into the width learning defect detection neural network model with incremental learning for training, and obtaining an optimized width learning defect detection neural network model with incremental learning; obtaining the image of the Mini-LED product to be detected, and performing defect detection on the image of the Mini-LED product to be detected by using the optimized width learning defect detection neural network model with incremental learning;

[0072] The method in the present invention has the advantages of simple structure, strong generalization ability and fast speed. The incremental learning improves the accuracy of defect detection by adding enhancement nodes. Compared with the deep neural network, the structure of the present invention is flat, the calculation is simple, and the training speed is relatively fast. The training of the width learning network does not require artificial prior parameter tuning, nor does it have a training process of slow gradient optimization until the objective function converges. The network performance is improved by forming enhancement nodes through feature mapping, and then the weight matrix can be obtained by solving the pseudo-inverse matrix, and finally the structure design of the width learning network is completed. In addition, during model training, there is no large amount of calculation, and the weights of the width learning network structure can be obtained only by simple matrix operations. Therefore, the efficiency and accuracy of defect detection and classification of Mini-LED products can be significantly improved. Description of the Drawings

[0073] Figure 1 It is a flowchart of a Mini-LED defect detection method based on width learning provided in Embodiment 1.

[0074] Figure 2 It is a structural diagram of a width learning defect detection neural network model with incremental learning provided in Embodiment 2.

[0075] Figure 3 It is a detection effect diagram of a Mini-LED defect detection method based on width learning provided in Embodiment 2.

[0076] Figure 4 It is a structural diagram of a Mini-LED defect detection system based on width learning provided in Embodiment 3. Detailed Embodiments

[0077] The drawings are only for illustrative purposes and should not be construed as a limitation of this patent;

[0078] To better illustrate this embodiment, some components in the drawings are omitted, enlarged or reduced, and do not represent the size of the actual product;

[0079] For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0080] The technical solutions of the present invention will be further described below with reference to the drawings and embodiments.

[0081] Embodiment 1

[0082] As Figure 1 shown, this embodiment provides a Mini-LED defect detection method based on width learning, including the following steps:

[0083] S1: Obtain the initial image of the Mini-LED product and perform preprocessing;

[0084] S2: Establish a width learning defect detection neural network model with incremental learning;

[0085] S3: Input the preprocessed initial image of the Mini-LED product into the width learning defect detection neural network model with incremental learning for training to obtain an optimized width learning defect detection neural network model with incremental learning;

[0086] S4: Obtain the image of the Mini-LED product to be detected, and use the optimized width learning defect detection neural network model with incremental learning to detect defects in the image of the Mini-LED product to be detected.

[0087] In the specific implementation process, first obtain the initial image of the Mini-LED product and perform preprocessing; establish a width learning defect detection neural network model with incremental learning; input the preprocessed initial image of the Mini-LED product into the width learning defect detection neural network model with incremental learning for training to obtain an optimized width learning defect detection neural network model with incremental learning; finally, obtain the image of the Mini-LED product to be detected, and use the optimized width learning defect detection neural network model with incremental learning to detect defects in the image of the Mini-LED product to be detected;

[0088] The method in the present invention has the advantages of simple structure, strong generalization ability and fast speed. The incremental learning improves the accuracy of defect detection by adding enhancement nodes; compared with the deep neural network, the structure of the present invention is flat, the calculation is simple, the training speed is relatively fast, the training of the width learning network does not require artificial prior parameter tuning, nor does it have a slow gradient optimization until the objective function converges. The network performance is improved by forming enhancement nodes through feature mapping, and then the weight matrix can be obtained by solving the pseudo-inverse matrix, and finally the width learning network structure design is completed; in addition, during model training, there is no large amount of calculation, and only simple matrix operations are required to obtain the weights of the width learning network structure, so the efficiency and accuracy of defect detection and classification of Mini-LED products can be significantly improved.

[0089] Embodiment 2

[0090] This embodiment provides a Mini-LED defect detection method based on width learning, including the following steps:

[0091] S1: Obtain the initial image of the Mini-LED product and perform preprocessing;

[0092] S2: Establish a width learning defect detection neural network model with incremental learning;

[0093] S3: Input the preprocessed initial image of the Mini-LED product into the width learning defect detection neural network model with incremental learning for training to obtain an optimized width learning defect detection neural network model with incremental learning;

[0094] S4: Obtain the image of the Mini-LED product to be detected, and use the optimized width learning defect detection neural network model with incremental learning to detect defects in the image of the Mini-LED product to be detected;

[0095] In the step S1, the specific method for obtaining and preprocessing the initial image of the Mini-LED product is as follows:

[0096] The types of the initial images of the Mini-LED products include initial images of normal products, initial images of products with foreign objects, and initial images of products with missing lamp beads;

[0097] The specific method for the preprocessing is as follows:

[0098] Divide all the initial images of the Mini-LED products into a training set and a test set, and perform grayscale operation and flattening operation on all the initial images of the Mini-LED products in sequence; fill the pixel values of all pixel points into a table, and each column of pixel values corresponds to an initial image of the Mini-LED product; mark different labels according to the different types of the initial images of the Mini-LED products, fill the labels into the first row of the corresponding table, and randomly shuffle the pixel information of each column of the table to complete the preprocessing of the initial images of the Mini-LED products;

[0099] The size of the initial image of the Mini-LED product is 200*215;

[0100] Mark different labels according to the different types of the initial images of the Mini-LED products, specifically:

[0101] Mark the initial image of the normal product as label 0, the initial image of the product with foreign objects as label 1, and the initial image of the product with missing lamp beads as label 2;

[0102] As Figure 2 shown, in the step S2, the width learning defect detection neural network model with incremental learning includes an input layer, a feature layer, an enhancement layer, and an output layer;

[0103] The output end of the input layer is connected to the input end of the feature layer, and the output end of the feature layer is connected to the input end of the enhancement layer;

[0104] The output end of the feature layer is also spliced with the output end of the enhancement layer and jointly connected to the input end of the output layer;

[0105] In step S3, the preprocessed initial image of the Mini-LED product is input into the width learning defect detection neural network model with incremental learning for training to obtain an optimized width learning defect detection neural network model with incremental learning. The specific method is as follows:

[0106] S3.1: Use the preprocessed initial image of the Mini-LED product as the input matrix X and input it into the input layer of the width learning defect detection neural network model with incremental learning for calculation to obtain the merged feature node matrix Z n =[Z1,Z2,…,Z n ∈R N×np ;

[0107] S3.2: Calculate the merged enhanced node matrix H according to the merged feature node matrix Z n =[Z1,Z2,…,Z n ∈R N×np Calculate the merged enhanced node matrix H m =[H1,H2,…,H m ;

[0108] S3.3: Merge the merged feature node matrix Z n =[Z1,Z2,…,Z n ∈R N×np with the merged enhanced node matrix H m =[H1,H2,…,H m again to obtain the first merged matrix, denoted as matrix A;

[0109] Matrix A satisfies A = [Z n |H m ;

[0110] S3.4: Add new enhanced nodes and merge them with matrix A to obtain the second merged matrix, denoted as matrix A m+1 ;

[0111] Matrix A m+1 satisfies A m+1 =[A|ξ(Z n W hm+1 +β hm+1 )];

[0112] S3.5: Calculate the pseudo-inverse matrix (A m+1 ) m+1 of matrix A + , and obtain the link weight W between matrix A m+1 and the output matrix Y according to the pseudo-inverse matrix (A m+1 ) + of matrix A m +1 ​m+1 ;

[0113] S3.6: Determine the link weight W m+1 between the judgment matrix A m+1 and the output matrix Y. If it meets the preset error condition, then at this time, the matrix A m+1 and the link weight W m+1 between it and the output matrix Y are saved to obtain an optimized width learning defect detection neural network model with incremental learning; otherwise, repeat steps S3.1 - S3.5;

[0114] In the said step S3.1, the specific method for obtaining the merged feature node matrix is:

[0115] Input the preset input matrix X into the input layer, and after n - times of feature mapping, obtain n groups of feature node matrices Z i , and the i - th group of feature node matrix Z i is specifically:

[0116] Z i = Φ(XW ei + β ei ) ∈ R N×p , i = 1, 2, …, n

[0117] where X is the input matrix, satisfying X ∈ R N×M , N represents the number of samples in the input dataset, and M represents the number of features of the input sample vector; W ei is the first weight matrix with a Gaussian distribution, satisfying W ei ∈ R M×p ; β ei is the first bias matrix, satisfying β ei ∈ R N×p ; p is the first parameter determined by the grid search method; the function Φ(·) represents a linear feature mapping;

[0118] Merge the n groups of feature node matrices Z i to obtain the merged feature node matrix, specifically:

[0119] Z n = [Z1, Z2, …, Z n ∈ R N×np

[0120] where Z n represents the merged feature node matrix;

[0121] In the said step S3.2, the specific method for obtaining the merged enhanced node matrix is:

[0122] Input the merged feature node matrix Z nAfter m times of feature mapping, m groups of enhanced node matrices H are obtained m , the m-th group of enhanced node matrix H m Specifically:

[0123] H m = ξ(Z n W hj + β hj ) ∈ R N×q , j = 1, 2, …, m

[0124] Among them, W hj is the second weight matrix with a Gaussian distribution, satisfying W hj ∈ R np×q ; β hj is the second bias matrix, satisfying β hj ∈ R N×q ; q is the second parameter determined by the grid search method; the function ξ(·) represents a non-linear feature mapping;

[0125] Merge the m groups of enhanced node matrices H m to obtain the merged enhanced node matrix. Specifically:

[0126] H m = [H1, H2, …, H m ∈ R N×mq

[0127] Among them, H m represents the merged enhanced node matrix;

[0128] In the step S3.5, the specific method for obtaining the link weight W m+1 between the matrix A m+1 and the output matrix Y is:

[0129] Merge the feature node matrix Z i of the feature layer and the enhanced node matrix H m of the enhanced layer to obtain the first merged matrix, denoted as matrix A. Specifically:

[0130] A = [Z n | H m ∈ R N×L , L = np + mq

[0131] The pseudo-inverse matrix A + of the matrix A is:

[0132]

[0133] Among them, λ is the regularization parameter, and I is the identity matrix;

[0134] According to the pseudo-inverse matrix A of the matrix A+ Obtain the link weight W between matrix A and matrix Y m , specifically:

[0135] W m =(A T A + λI) -1 A T Y

[0136] Obtain a new enhanced node matrix, merge the new enhanced node matrix with matrix A to obtain a second merged matrix, denoted as matrix A m+1 , specifically:

[0137] A m+1 =[A|ξ(Z n W hm+1 +β hm+1 )]

[0138] where W hm+1 is the third weight matrix with a Gaussian distribution, satisfying W hm+1 ∈R np×k ; β hm+1 is the third bias matrix, satisfying β hm+1 ∈R k , and k is the third parameter determined by the grid search method;

[0139] The pseudo-inverse matrix (A m+1 ) m+1 ) + is:

[0140]

[0141] where D = A + ξ(Z n W hm+1 +β hm+1 ); C = ξ(Z n W hm+1 +β hm+1 ) - AD;

[0142] According to the pseudo-inverse matrix (A m+1 ) m+1 ) + obtain the link weight W m+1 between matrix A m+1 and the output matrix Y

[0143]

[0144] where W m+1 is the link weight between matrix A m+1 and the output matrix Y.

[0145] In the specific implementation process, first, obtain the initial image of the Mini-LED product and perform preprocessing;

[0146] In this embodiment, the size of the initial image of the Mini-LED product is 200*215, and the types of the initial image of the Mini-LED product include the initial image of a normal product, the initial image of a foreign object product, and the initial image of a missing lamp bead product;

[0147] Divide all the initial images of the Mini-LED products into a training set and a test set. Perform grayscale conversion on all the initial images of the Mini-LED products, that is, convert the color pictures into grayscale pictures between 0 and 255. Then perform a flattening operation on the converted grayscale pictures to pull the picture data into a one-dimensional vector. Then fill the one-dimensional vector into a table to form a 43000*1000 table. Each column of pixel values corresponds to an initial image of a Mini-LED product; Mark different labels according to the different types of the initial images of the Mini-LED products, and fill the labels into the first row of the corresponding table. The initial image of a normal product is marked as label 0, the initial image of a foreign object product is marked as label 1, and the initial image of a missing lamp bead product is marked as label 2;

[0148] Randomly shuffle the pixel information of each column of the table to prevent overfitting in model training, and complete the preprocessing of the initial images of the Mini-LED products;

[0149] After that, establish a width learning defect detection neural network model with incremental learning; Input the preprocessed initial images of the Mini-LED products into the width learning defect detection neural network model with incremental learning for training to obtain an optimized width learning defect detection neural network model with incremental learning. The specific method is as follows:

[0150] S3.1: Use the preprocessed initial image of the Mini-LED product as the input matrix X and input it into the input layer of the width learning defect detection neural network model with incremental learning for calculation to obtain the merged feature node matrix Z n =[Z1,Z2,…,Z n ∈R N×np ;

[0151] S3.2: Calculate and obtain the merged enhanced node matrix H according to the merged feature node matrix Z n =[Z1,Z2,…,Z n ∈R N×np Calculate and obtain the merged enhanced node matrix H m =[H1,H2,…,H m ;

[0152] S3.3: Use the merged feature node matrix Z n= [Z1, Z2, …, Z n ∈ R N×np is merged with the combined enhanced node matrix H m = [H1, H2, …, H m again to obtain the first merged matrix, denoted as matrix A;

[0153] Matrix A satisfies A = [Z n |H m ;

[0154] S3.4: Add new enhanced nodes and merge them with matrix A to obtain the second merged matrix, denoted as matrix A m+1 ;

[0155] Matrix A m+1 satisfies A m+1 = [A|ξ(Z n W hm+1 + hm+1 )];

[0156] S3.5: Calculate the pseudo-inverse matrix of matrix A m+1 ( m+1 ) + , and obtain the link weight W m+1 between matrix A m+1 and the output matrix Y according to the pseudo-inverse matrix of matrix A + ; m+1 m+1 m+1 ;

[0157] S3.6: Determine whether the link weight W m+1 between matrix A m+1 and the output matrix Y meets the preset error condition. If it meets, save the link weight W m+1 between the current matrix A m+1 and the output matrix Y to obtain an optimized width learning defect detection neural network model with incremental learning; otherwise, repeat steps S3.1 - S3.5;

[0158] Finally, obtain the image of the Mini-LED product to be detected, and use the optimized width learning defect detection neural network model with incremental learning to detect the defects in the image of the Mini-LED product to be detected;

[0159] In this embodiment, W ei , β ei , W hj , β hj , W hm+1 and β hm+1 are all randomly generated by the sparse autoencoder, aiming to better sparsely represent the input matrix;

[0160] Figure 3 The figure shows the effect diagram of using the method in this embodiment to detect the appearance defects of Mini-LED products. It can be seen from the figure that the detection effect is good and the accuracy is high;

[0161] The method in the present invention has the advantages of simple structure, strong generalization ability and fast speed. The incremental learning improves the accuracy of defect detection by adding enhancement nodes. Compared with the deep neural network, the structure of the present invention is flat, the calculation is simple, and the training speed is relatively fast. The training of the width learning network does not require manual pre-parameter adjustment, nor does it have a slow gradient optimization process until the objective function converges. The network performance is improved by forming enhancement nodes through feature mapping, and then the weight matrix can be obtained by solving the pseudo-inverse matrix, finally completing the design of the width learning network structure. In addition, during model training, there is no large amount of calculation, and the weights of the width learning network structure can be obtained only by simple matrix operations. Therefore, the efficiency and accuracy of defect detection and classification of Mini-LED products can be significantly improved.

[0162] Embodiment 3

[0163] As Figure 4 shown, this embodiment provides a Mini-LED defect detection system based on width learning, which applies the Mini-LED defect detection method based on width learning described in Embodiment 1 or 2, including:

[0164] A preprocessing unit 301: used to obtain the initial image of the Mini-LED product and perform preprocessing;

[0165] A model construction unit 302: used to establish a width learning defect detection neural network model with incremental learning;

[0166] A model training unit 303: used to input the preprocessed initial image of the Mini-LED product into the width learning defect detection neural network model with incremental learning for training, and obtain an optimized width learning defect detection neural network model with incremental learning;

[0167] A defect prediction unit 304: used to obtain the image of the Mini-LED product to be detected, and perform defect detection on the image of the Mini-LED product to be detected by using the optimized width learning defect detection neural network model with incremental learning.

[0168] In the specific implementation process, first, the preprocessing unit 301 acquires the initial image of the Mini-LED product and performs preprocessing; the model construction unit 302 establishes a width learning defect detection neural network model with incremental learning; the model training unit 303 inputs the preprocessed initial image of the Mini-LED product into the width learning defect detection neural network model with incremental learning for training to obtain an optimized width learning defect detection neural network model with incremental learning; finally, the defect prediction unit 304 acquires the image of the Mini-LED product to be detected and uses the optimized width learning defect detection neural network model with incremental learning to perform defect detection on the image of the Mini-LED product to be detected;

[0169] The method in the present invention has the advantages of simple structure, strong generalization ability and fast speed. Incremental learning improves the accuracy of defect detection by adding enhancement nodes; compared with the deep neural network, the structure of the present invention is flat, the calculation is simple, and the training speed is relatively fast. The training of the width learning network does not require manual pre-parameter adjustment, nor does it have a training process of slow gradient optimization until the objective function converges. The network performance is improved by forming enhancement nodes through feature mapping, and then the weight matrix can be obtained by solving the pseudo-inverse matrix, and finally the structure design of the width learning network is completed; in addition, during model training, there is no large amount of calculation, and only simple matrix operations are required to obtain the weights of the width learning network structure. Therefore, the efficiency and accuracy of defect detection and classification of Mini-LED products can be significantly improved.

[0170] The same or similar reference numerals correspond to the same or similar components;

[0171] The terms describing the positional relationship in the drawings are only for illustrative purposes and cannot be construed as a limitation of this patent;

[0172] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A Mini-LED defect detection method based on width learning, characterized in that, It includes the following steps: S1: Obtain the initial image of the Mini-LED product and perform preprocessing; S2: Establish a width learning defect detection neural network model with incremental learning; S3: Input the preprocessed initial image of the Mini-LED product into the width learning defect detection neural network model with incremental learning for training to obtain an optimized width learning defect detection neural network model with incremental learning; specifically: S3.1: Use the preprocessed initial image of the Mini-LED product as the input matrix Input it into the input layer of the width learning defect detection neural network model with incremental learning for calculation to obtain the merged feature node matrix ; where is the number of feature node matrices; represents the number of samples in the input dataset; is the first parameter determined by the grid search method S3.2: According to the merged feature node matrix calculate and obtain the merged enhanced node matrix ; where is the number of enhanced node matrices S3.3: The merged feature node matrix and the merged enhanced node matrix are merged again to obtain a first merged matrix, denoted as matrix ; Matrix Satisfy ; S3.4: Add a new enhanced node and merge it with the matrix to obtain a second merged matrix, denoted as matrix ; Matrix satisfies ; wherein is the third weight matrix with a Gaussian distribution and satisfies ; is the third bias matrix and satisfies , is the third parameter determined by the grid search method; the function represents a non - linear feature mapping; S3.5: Calculate the matrix of the pseudo-inverse matrix , and according to the pseudo-inverse matrix of the matrix obtain the link weight between the matrix and the output matrix ; ​ S3.6: Judgment matrix and the output matrix whether the link weight between them meets the preset error condition. If it meets, then the matrix at this time and the output matrix the link weight between them is saved to obtain an optimized width learning defect detection neural network model with incremental learning; otherwise, repeat steps S3.1~S3.5; S4: Obtain the image of the Mini-LED product to be detected, and use the optimized width learning defect detection neural network model with incremental learning to detect defects in the image of the Mini-LED product to be detected.

2. The method for defect detection of Mini-LED based on width learning according to claim 1, wherein, In the step S1, the specific method for obtaining the initial image of the Mini-LED product and performing preprocessing is: The types of the initial images of the Mini-LED products include initial images of normal products, initial images of products with foreign objects, and initial images of products with missing lamp beads; The specific method for the preprocessing is: Divide all the initial images of the Mini-LED products into a training set and a test set, and perform grayscale operation and flattening operation on all the initial images of the Mini-LED products in sequence; fill the pixel values of all pixel points into a table, and each column of pixel values corresponds to an initial image of a Mini-LED product; mark different labels according to the different types of the initial images of the Mini-LED products, fill the labels into the first row of the corresponding table, and perform a random shuffling operation on the pixel information of each column of the table to complete the preprocessing of the initial images of the Mini-LED products.

3. The method for defect detection of Mini-LED based on width learning according to claim 2, wherein, The size of the initial image of the Mini-LED product is 200*215.

4. A Mini-LED defect detection method based on width learning according to claim 3, characterized in that Mark different labels according to the different types of the initial images of the Mini-LED products, specifically: Mark the initial image of the normal product as label 0, mark the initial image of the product with foreign objects as label 1, and mark the initial image of the product with missing lamp beads as label 2.

5. A Mini-LED defect detection method based on width learning according to claim 4, characterized in that, In the step S2, the width learning defect detection neural network model with incremental learning includes an input layer, a feature layer, an enhancement layer, and an output layer; The output end of the input layer is connected to the input end of the feature layer, and the output end of the feature layer is connected to the input end of the enhancement layer; The output end of the feature layer is also spliced with the output end of the enhancement layer and jointly connected to the input end of the output layer.

6. A Mini-LED defect detection method based on width learning according to claim 1, characterized in that In the step S3.1, the specific method for obtaining the merged feature node matrix is: Input the preset input matrix into the input layer, and after times of feature mapping, obtain groups of feature node matrices . The i-th group of feature node matrices is specifically as follows: Among them, is the input matrix, satisfying , represents the number of samples in the input dataset, represents the number of features of the input sample vector; is the first weight matrix with a Gaussian distribution, satisfying ; is the first bias matrix, satisfying ; is the first parameter determined by the grid search method; the function represents a linear feature mapping; Merge group feature node matrix to obtain the merged feature node matrix, specifically as follows: Among them, represents the merged feature node matrix.

7. A Mini-LED defect detection method based on width learning according to claim 6, characterized in that In the step S3.2, the specific method for obtaining the merged enhancement node matrix is: The merged feature node matrix obtained after m times of feature mapping group of enhanced node matrices , the m-th group of enhanced node matrices Specifically: Among them, is the second weight matrix with a Gaussian distribution, satisfying ; is the second bias matrix, satisfying ; is the second parameter determined by the grid search method; the function represents a non-linear feature mapping; Merge group of enhanced node matrices to obtain the merged enhanced node matrix, specifically: Among them, represents the merged enhanced node matrix.

8. A Mini-LED defect detection method based on width learning according to claim 7, characterized in that, In the step S3.5, obtain the matrix and the output matrix The specific method for the link weight between them is as follows: Merge the feature node matrix of the feature layer with the enhanced node matrix of the enhancement layer to obtain a first merged matrix, denoted as matrix , specifically: Matrix pseudo-inverse matrix is as follows: Among them, is the regularization parameter, is the identity matrix; According to the matrix pseudo-inverse matrix to obtain the matrix and the matrix link weight between , specifically: Obtain a new enhanced node matrix, and merge the new enhanced node matrix with matrix to obtain a second merged matrix, denoted as matrix , specifically: Among them, is the third weight matrix with a Gaussian distribution, satisfying ; is the third bias matrix, satisfying , is the third parameter determined by the grid search method; Matrix pseudo-inverse matrix is as follows: Among them, ; , ; According to the matrix pseudo-inverse matrix of to obtain the matrix and the output matrix link weights between , specifically: Among them, is the matrix and the output matrix is the link weight therebetween.

9. A Mini-LED defect detection system based on width learning, which applies a Mini-LED defect detection method described in any one of claims 1 to 8, is characterized in that It includes: A preprocessing unit: used to obtain the initial image of the Mini-LED product and perform preprocessing; A model construction unit: used to establish a width learning defect detection neural network model with incremental learning; A model training unit: used to input the preprocessed initial image of the Mini-LED product into the width learning defect detection neural network model with incremental learning for training to obtain an optimized width learning defect detection neural network model with incremental learning; Defect prediction unit: used to obtain the Mini-LED product image to be detected, and perform defect detection on the Mini-LED product image to be detected by using the optimized width learning defect detection neural network model with incremental learning.

Images