Substrate defect detection method based on incremental learning CNN tree model

Abstract

The invention discloses a substrate defect detection method based on an incremental learning CNN tree model, and the method comprises the following steps: constructing a source node network, enabling multiple layers of nodes to form an initial node network in a parallel cascading manner, and enabling each node to be provided with a substrate defect classification CNN network; calculating a three-dimensional likelihood matrix of the input substrate image, performing descending sorting on value elements of the three-dimensional likelihood matrix, and establishing a corresponding Softmax likelihood value table for each sub-node; adding a defect class to be detected to sub-nodes of the initial node network by using a semi-supervised growth rule, and adapting to a new substrate defect sample class or a new-old combined substrate defect sample class in a tree-shaped growth mode; retraining or finely adjusting a hidden layer of the substrate defect classification CNN network based on a back propagation strategy; and based on the substrate defect classification CNN network of each node of the trained tree network, outputting a substrate defect type result. According to the invention, accurate and rapid substrate defect detection can be realized, and the robustness is better.

Description

Technical Field

[0001] The present invention relates to the technical field of substrate defect detection, and in particular to a substrate defect detection method based on an incremental learning CNN tree model. Background Art

[0002] In microelectronics manufacturing, surface defects of flexible substrates are important factors that directly affect the final performance of packaging substrates. Timely and accurate selection of defective substrates is the most critical link to avoid huge economic losses to electronic processing manufacturers and effectively save costs. Packaging substrates are increasingly highly integrated and multi-layer component packaging to meet people's increasing demand for the functions of electronic products. The corresponding substrate manufacturing technology and processing technology have become more and more complex, which has led to a significant increase in the risk of substrate failure and the difficulty of defect detection. At present, there are two main methods for dealing with the problem of qualified substrate inspection. One is manual visual sampling. Due to the subjective decision-making and fatigue factors of quality inspectors, visual inspection has problems such as high misjudgment rate and low efficiency; the other is online discrimination based on intelligent detection algorithms. These methods use precision mechanical positioning mechanisms combined with intelligent control algorithms to improve the efficiency and accuracy of substrate defect detection.

[0003] The development of intelligent defect detection algorithms is an inevitable trend of automatic identification systems. As the types of surface defects on flexible substrates are becoming increasingly complex, the traditional methods based on manual feature extraction and analysis have limited defect detection efficiency and accuracy. Therefore, the introduction and design of novel CNN network models have become a research hotspot in the field of intelligent detection of flexible substrate defects. After the development in recent years, many substrate detection models with superior performance have been produced in China. However, in actual semiconductor packaging and testing projects, it is usually difficult to collect a large number of comprehensive and balanced defect samples, which makes the substrate defect detection method based on a single CNN or static hierarchical cascade CNN network model have a high defect omission and misjudgment rate under the challenge of small sample drive and unbalanced data set; at the same time, the existing defect detection model is difficult to handle the substrate defect recognition under the dynamic addition of new defect samples in the production process. Therefore, it is an urgent need to design a new defect intelligent detection model based on CNN to achieve detection performance that meets production needs under the challenge of specific sample patterns, which has important application value in the actual semiconductor packaging and testing field. Summary of the invention

[0004] In order to overcome the defects and shortcomings of the prior art, the present invention provides a substrate defect detection method based on an incremental learning CNN tree model. The present invention solves the problem of limited substrate defect detection performance of various models based on a single CNN under the challenges of small sample drive, unbalanced sample quantity and the addition of new defective samples in the semiconductor packaging and testing process. The present invention can achieve accurate and fast substrate defect detection, has better robustness, and is easier to expand to apply to more types of substrate surface defect detection.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention provides a substrate defect detection method based on an incremental learning CNN tree model, comprising the following steps:

[0007] Constructing a source node network, forming an initial node network by parallel cascading multiple layers of nodes, each node is provided with a substrate defect classification CNN network, the substrate defect classification CNN network includes an input layer, multiple convolutional layers, a fully connected layer and a Softmax output layer;

[0008] Through the Softmax output layer of the substrate defect classification CNN network, the three-dimensional likelihood matrix of the input substrate image is calculated respectively, the value elements of the three-dimensional likelihood matrix are sorted in descending order, and the corresponding Softmax likelihood value table is established for each child node;

[0009] Based on the sorted three-dimensional likelihood matrix, the defect classes to be detected are added to the child nodes of the initial node network using a semi-supervised growth rule. The initial network nodes start from a single root node and adapt to the new or new and old combination of substrate defect sample classes in a tree-like growth manner to achieve orderly classification of substrate defects.

[0010] Based on the back propagation strategy of different network layer depths, the hidden layer of the substrate defect classification CNN network is retrained or fine-tuned using the newly added substrate defects and the initial defect data set to obtain the substrate defect classification CNN network at each node of the trained tree network;

[0011] Based on the substrate defect classification CNN network at each node of the trained tree network, the substrate defect type results are output.

[0012] As a preferred technical solution, the substrate defect classification CNN network adopts a CNN network framework based on a dilated spatial convolution pooling pyramid and an ECA attention mechanism module. The dilated spatial convolution pooling pyramid stacks dilated convolutions with different dilated rates in parallel or in cascade to obtain multi-scale substrate defect information gain. The ECA attention mechanism module adaptively adjusts the weights between channels.

[0013] As a preferred technical solution, the Softmax output layer converts the response of CNN to the image into an exponential scale by using a Softmax function instead of the number of images classified into each sub-node, thereby obtaining a three-dimensional likelihood matrix of the input image.

[0014] As a preferred technical solution, the Softmax function is expressed as:

[0015]

[0016] Among them, y represents the output nonlinear probability, W i represents the weight matrix between the output layer and the i-th neuron in the previous layer, B is the bias vector, S1 is the output of the fully connected layer of the network, and k represents the number of output types.

[0017] As a preferred technical solution, the three-dimensional likelihood matrix is ​​expressed as:

[0018] S={S [i] |S [i] .label,S [i] .value = [y1, y2, y3], S [i] .nodes = [n1, n2, n3]}

[0019] Among them, S is the output three-dimensional likelihood matrix, S [i] represents the component of the i-th three-dimensional likelihood matrix S, S [i] .label represents the predicted label of the input image, S [i] .value represents the likelihood value vector, S [i] .nodes is the current node number.

[0020] As a preferred technical solution, based on the sorted three-dimensional likelihood matrix, the defect class to be detected is added to the child nodes of the initial node network using a semi-supervised growth rule. The initial network node starts from a single root node and adapts to the new or new and old combination of substrate defect sample classes in a tree-like growth manner, specifically including:

[0021] Set the depth D and width constraint W of the maximum child node of the tree network, and set the inductive bias of the child node; the semi-supervised growth rule is expressed as:

[0022]

[0023] Among them, y1, y2, y3 are the Softmax output values ​​of the three-dimensional likelihood matrix, {C|C1,…,C k} represents a set C of substrate defect types with element size k, {children|children1,…,children k} represents the child node set children whose element size is i;

[0024] According to the three-dimensional likelihood matrix of the input substrate image, combined with the Softmax likelihood value table, a table lookup operation is performed, and the input substrate defect class is added to the child node of the initial node network to achieve orderly defect classification;

[0025] If y1-y2>β and y1>y2, then add the current new substrate defect class to the corresponding child node;

[0026] If y1-y2<β and y1-y3>α, the current new substrate defect class is combined into a new child node;

[0027] If y1-y2<β and y1-y3<α, the current new substrate defect class is added as a new child node, and the source node network is horizontally expanded, where β and α are thresholds selected by the learning algorithm.

[0028] As a preferred technical solution, the depth D and width constraint W of the maximum child node of the tree network are specifically expressed as:

[0029]

[0030] Among them, D max Indicates the preset maximum depth of a positive integer N + |children i | represents the number of child nodes on the same layer under node i in the tree network;

[0031] And / or, setting the inductive preference of the tree network nodes based on the Occam's razor principle.

[0032] As a preferred technical solution, based on the back propagation strategy of different network layer depths, the hidden layer of the substrate defect classification CNN network is retrained or fine-tuned using the newly added substrate defects and initial defect data sets, specifically including:

[0033] Add the substrate defect class to be added to the sub-node of the currently working sub-node CNN network, and reorganize the newly added substrate defects and initial training data;

[0034] According to the three-dimensional likelihood matrix information of the newly added nodes, the hidden layer of the substrate defect classification CNN network is retrained or fine-tuned using a back propagation strategy based on different network depths;

[0035] After the local CNN network parameters in the tree network are updated, the tree network enters the next batch of substrate defect classification and local CNN network model iterative update until no new substrate defect types are detected. The CNN network of each node of the tree network stops iterative update, and finally the substrate defect classification CNN network on each node of the trained tree network is obtained.

[0036] As a preferred technical solution, the back propagation strategy is expressed as:

[0037]

[0038] Among them, θ represents the parameters of the network, η is the learning rate, J represents the loss function, and x (i:i+n) Represents the characteristic values ​​of the i-th to i+n-th training samples, y (i:i+n) denotes the labels of the training samples of corresponding size, Represents the gradient of solving the network parameter θ.

[0039] As a preferred technical solution, a Dropout layer is introduced to enhance regularization during the process of retraining or fine-tuning the hidden layer of the substrate defect classification CNN network, which is specifically expressed as follows:

[0040]

[0041] Among them, Π represents the indicator function, r i is a random variable that follows the Bernoulli distribution Bernoulli(p), x i is the input of the i-th neuron, y i is the output of the i-th neuron after Dropout processing.

[0042] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0043] (1) The present invention adopts a substrate surface defect detection tree model based on hierarchical cascaded CNN, which provides a new detection method for complex defects on the surface of packaged substrates during semiconductor packaging and testing. The tree model of the present invention only iteratively trains the CNN network on local nodes during each learning process, thereby improving the efficiency of substrate defect detection; in the tree model, the baseline network will only be activated when the parent node classifies the input of the current branch node, thereby improving the efficiency of substrate defect detection.

[0044] (2) The present invention adopts an incremental learning strategy to enable the nodes of the tree model to self-grow in an orderly manner. The tree model locally self-learns and iterates as new data emerges, which solves the problem of defect detection when the existing model cannot cope with the challenge of adding new defect types and improves the robustness of the model. In the iterative learning process of the tree model, the combination of new and old samples is used to improve the learning performance of the CNN network model when it is difficult to obtain a large number of defect samples with a balanced number, thereby improving the accuracy of the tree model in detecting defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of a substrate defect detection method based on an incremental learning CNN tree model of the present invention;

[0046] Figure 2 This is the overall architecture diagram of the incremental learning CNN tree model of the present invention;

[0047] Figure 3 A schematic diagram of the network framework of the CNN model for substrate defect classification of the present invention;

[0048] Figure 4 Schematic diagram of the incremental learning phase of the CNN tree model of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0050] Example

[0051] like Figure 1 As shown, this embodiment provides a substrate defect detection method based on incremental learning CNN tree model, comprising the following steps:

[0052] S1: Figure 2 As shown, an incremental learning CNN tree model is constructed, which specifically includes:

[0053] Source node network: The initial node network is composed of three layers of nodes in parallel cascade mode. The types of nodes are defined as root nodes, branch nodes and leaf nodes. Each parent node is composed of a classified CNN homogeneous base algorithm. The classified CNN homogeneous base algorithm refers to the classified CNN network on each node in the source node network.

[0054] Substrate defect classification CNN network: including the CNN network framework that introduces the Atrous Spatial Convolution Pooling Pyramid (ASPP) and ECA attention mechanism modules, and uses the Softmax function to convert the output layer's response to the image into an exponential scale to replace the number of images classified into each child node;

[0055] Considering that the input flexible substrate image has multiple granularities and the substrate surface defects have a large aspect ratio, the atrous spatial convolution pooling pyramid (ASPP) stacks atrous convolutions with different atrous rates in parallel or in cascade to obtain multi-scale substrate defect information gain. The convolution kernel size after m convolutions are superimposed is expressed as:

[0056]

[0057] Among them, K is the original convolution kernel size, d is the dilation rate, K m,d Represents the receptive field size of m convolution kernels with d dilation rate.

[0058] In view of the high resolution of the input flexible substrate image and the presence of complex background redundant information, the ECA attention mechanism module distinguishes the importance of each channel feature by adaptively adjusting the weights between channels to improve the feature expression ability of the model. The kernel size involved in the ECA attention mechanism module is calculated as:

[0059]

[0060] Among them, K e is the size of the convolution kernel, in_channel represents the number of input channels, γ and b are adjustable parameters, γ is a hyperparameter selected based on experience and is usually set to 2;

[0061] Source node network growth mechanism based on incremental learning: including the self-growth rule of the source node network based on semi-supervision, which can make the initial network node grow in a tree-like manner from a single root node to adapt to new or new and old combinations of substrate defect sample classes, and at the same time, it can iteratively train the base algorithm to obtain a CNN network model with better performance;

[0062] like Figure 3 The figure shows the network framework of the CNN model for substrate defect classification. The CNN network framework includes 1 input layer, 6 convolutional layers, 1 fully connected layer and 1 output Softmax layer. The specific parameters are as follows:

[0063]

[0064] Among them, K w ,K h , N represents the width, height and number of convolution kernels, S w ,S h They represent the width and height of the average pooling kernel AvgPooling respectively, and ReLU is the activation function.

[0065] Compared with the current substrate surface defect detection models based on various single CNN network frameworks, the tree model of this embodiment adopts an improved CNN network framework and an expandable mechanism of the source node network based on incremental learning to achieve more robust and accurate substrate defect detection under specific sample challenges.

[0066] S2: Calculate the three-dimensional likelihood matrix of the input substrate image through the CNN network based on the Softmax output layer function on each child node in the three-layer source node network, and sort the value elements of the three-dimensional likelihood matrix in descending order, specifically including:

[0067] S21: The input substrate image is processed using the CNN network on each child node in the three-layer source node network, including image preprocessing, convolution, maximum pooling and other operations; the output layer Softmax function converts the CNN response to the image into an exponential scale to replace the number of images classified into each child node, and obtains the three-dimensional likelihood matrix of the input image;

[0068] In this embodiment, the Softmax function of the output layer of the CNN network is expressed as:

[0069]

[0070] Among them, y represents the output nonlinear probability, W i represents the weight matrix between the output layer and the i-th neuron in the previous layer, B is the bias vector, S1 is the output of the fully connected layer of the network, and k represents the number of output types;

[0071] S22: sorting the element values ​​of the three-dimensional likelihood matrix of the input substrate image in descending order, and establishing a corresponding Softmax likelihood value table for each child node;

[0072] In this embodiment, the three-dimensional likelihood matrix of the input substrate image is expressed as:

[0073] S={S [i] |S [i] .label,S [i] .value = [y1, y2, y3], S [i] .nodes = [n1, n2, n3]}

[0074] Among them, S is the output three-dimensional likelihood matrix, S [i] represents the component of the i-th three-dimensional likelihood matrix S, S [i] .label represents the predicted label of the input image, S [i] .value represents the likelihood value vector, S [i] .nodes is the current node number;

[0075] In this embodiment, the number of layers and nodes of the source node network is determined by the number of substrate defect types in the initial training data set.

[0076] S3: Figure 4 As shown in the figure, based on the sorted three-dimensional likelihood matrix, the defect class to be detected is added to the child node of the three-layer source node network using the semi-supervised growth rule to achieve orderly classification of substrate defects, including:

[0077] S31: Set the depth D and width constraint W of the maximum child node of the tree network model, and set the inductive bias of the child node;

[0078] In this embodiment, the depth D and width constraint W of the maximum child node of the tree network model are expressed as:

[0079]

[0080] Among them, D max Indicates the preset maximum depth of a positive integer N + |children i | represents the number of child nodes on the same layer under node i in the tree model;

[0081] In this embodiment, the inductive preference of the nodes of the tree network model is designed based on the principle of Occam's razor to prevent the nodes from falling into local cycle expansion and avoid classification failure of the tree network model during the node expansion process;

[0082] In this embodiment, the semi-supervised growth rule is expressed as:

[0083]

[0084] Among them, y1, y2, y3 are the Softmax output values ​​of the three-dimensional likelihood matrix, {C|C1,…,C k} represents a set C of substrate defect types with element size k, {children|children1,…,children k} represents the child node set children whose element size is i.

[0085] S32: Based on the three-dimensional likelihood matrix of the input substrate image, a table lookup operation is performed in combination with the Softmax likelihood value table, and the input substrate defect class is added to the child node of the three-layer source node network to achieve orderly defect classification.

[0086] S33: If y1-y2>β and y1>y2, it means that the new substrate defect class has a strong correlation with a specific child node of the source node network, and the class will be added to the corresponding child node;

[0087] S34: If y1-y2<β and y1-y3>α, it means that the new substrate defect class is likely to exceed one child node, and the class will be combined into a new child node;

[0088] S35: If y1-y2<β and y1-y3<α, it means that the likelihood value of the new substrate defect class is not significantly greater than other values ​​or all child nodes are full, and the new substrate defect class will be added as a new child node to horizontally expand the source node network. β and α in this embodiment are thresholds selected by the learning algorithm;

[0089] S4: According to the different depths of the baseline network, the hidden layer of the substrate defect classification CNN network is retrained or fine-tuned based on the back propagation strategy of different network layer depths using the newly added substrate defects and initial defect datasets to obtain a CNN-based algorithm with better performance, including:

[0090] S41: After the substrate defect class is added to the child node of the corresponding baseline network, the newly added substrate defect and the initial training data are reorganized;

[0091] In this embodiment, the baseline network is the currently working sub-node CNN network in the tree model.

[0092] S42: According to the three-dimensional likelihood matrix information of the newly added nodes, the hidden layer of the substrate defect classification CNN network is retrained or fine-tuned using a back-propagation strategy based on different network depths to obtain a tree model-based algorithm with better performance.

[0093] In this embodiment, the back propagation strategy is designed based on a small batch gradient descent strategy, which is expressed as:

[0094]

[0095] Among them, θ represents the parameters of the network, η (0<η<1) is the learning rate, J represents the loss function, and x (i:i+n) Represents the characteristic values ​​of the i-th to i+n-th training samples, y (i:i+n) denotes the labels of the training samples of corresponding size, Represents the gradient of solving the network parameter θ;

[0096] In this embodiment, as the tree network model goes deeper, the data set size of the branch nodes will shrink, resulting in a higher chance of overfitting for the branch nodes than for the root nodes. More Dropout layers are introduced in the process of retraining or fine-tuning the CNN network to enhance regularization:

[0097]

[0098] Among them, Π represents the indicator function, r i is a random variable that follows the Bernoulli distribution Bernoulli(p), x i is the input of the i-th neuron, y i is the output of the i-th neuron after Dropout processing;

[0099] S43: After the local CNN network parameters in the tree model are updated, the tree model enters the next batch of substrate defect classification and local CNN network model iterative update; until no new substrate defect types are detected, the CNN network of each node of the tree model stops iterative updating, and finally the CNN network model on each node of the trained tree model is obtained.

[0100] Due to the lack of training data sets in the actual substrate industry, detection is first performed based on learning with a small number of samples. The detected defects are added to the training data set for retraining and then retested. The above implementation steps are a process of iterative cycles and orderly growth of nodes, which can be easily expanded to adapt to more types of substrate defect detection.

[0101] S5: Based on the substrate defect classification CNN network at each node of the trained tree network, the substrate defect type results are output.

[0102] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A substrate defect detection method based on incremental learning CNN tree model, characterized in that: The steps include: Constructing a source node network, forming an initial node network by parallel cascading multiple layers of nodes, each node is provided with a substrate defect classification CNN network, the substrate defect classification CNN network includes an input layer, multiple convolutional layers, a fully connected layer and a Softmax output layer; Through the Softmax output layer of the substrate defect classification CNN network, the three-dimensional likelihood matrix of the input substrate image is calculated respectively, the value elements of the three-dimensional likelihood matrix are sorted in descending order, and the corresponding Softmax likelihood value table is established for each child node; Based on the sorted three-dimensional likelihood matrix, the defect classes to be detected are added to the child nodes of the initial node network using a semi-supervised growth rule. The initial network nodes start from a single root node and adapt to the new or new and old combination of substrate defect sample classes in a tree-like growth manner to achieve orderly classification of substrate defects. Based on the back propagation strategy of different network layer depths, the hidden layer of the substrate defect classification CNN network is retrained or fine-tuned using the newly added substrate defects and the initial defect data set to obtain the substrate defect classification CNN network at each node of the trained tree network; Based on the substrate defect classification CNN network at each node of the trained tree network, the substrate defect type results are output.

2. The substrate defect detection method based on incremental learning CNN tree model according to claim 1 is characterized in that: The substrate defect classification CNN network adopts a CNN network framework based on a dilated spatial convolution pooling pyramid and an ECA attention mechanism module. The dilated spatial convolution pooling pyramid stacks dilated convolutions with different dilated rates in parallel or in cascade to obtain multi-scale substrate defect information gain. The ECA attention mechanism module adaptively adjusts the weights between channels.

3. The substrate defect detection method based on incremental learning CNN tree model according to claim 1 is characterized in that: The Softmax output layer converts the response of CNN to the image into an exponential scale by using the Softmax function instead of the number of images classified into each sub-node, thereby obtaining a three-dimensional likelihood matrix of the input image.

4. The substrate defect detection method based on incremental learning CNN tree model according to claim 3 is characterized in that: The Softmax function is expressed as: Among them, y represents the output nonlinear probability, W i represents the weight matrix between the output layer and the i-th neuron in the previous layer, B is the bias vector, S1 is the output of the fully connected layer of the network, and k represents the number of output types.

5. The substrate defect detection method based on incremental learning CNN tree model according to claim 1 is characterized in that: The three-dimensional likelihood matrix is ​​expressed as: S={S [i] |S [i] .label,S [i] .value=[y1,y2,y3],S [i] .nodes=[n1,n2,n3]} Among them, S is the output three-dimensional likelihood matrix, S [i] represents the component of the i-th three-dimensional likelihood matrix S, S [i] .label represents the predicted label of the input image, S [i] .value represents the likelihood value vector, S [i] .nodes is the current node number.

6. The substrate defect detection method based on incremental learning CNN tree model according to claim 1 is characterized in that: Based on the sorted three-dimensional likelihood matrix, the defect classes to be detected are added to the child nodes of the initial node network using a semi-supervised growth rule. The initial network nodes start from a single root node and grow in a tree-like manner to adapt to new or new and old combinations of substrate defect sample classes, including: Set the depth D and width constraint W of the maximum child node of the tree network, and set the inductive bias of the child node; The semi-supervised growth rule is expressed as: Among them, y1, y2, y3 are the Softmax output values ​​of the three-dimensional likelihood matrix, {C|C1,…,C k } represents a set C of substrate defect types with element size k, {children|children1,...,children k } represents the child node set children whose element size is i; According to the three-dimensional likelihood matrix of the input substrate image, combined with the Softmax likelihood value table, a table lookup operation is performed, and the input substrate defect class is added to the child node of the initial node network to achieve orderly defect classification; If y1-y2>β and y1>y2, then add the current new substrate defect class to the corresponding child node; If y1-y2<β and y1-y3>α, the current new substrate defect class is combined into a new child node; If y1-y2<β and y1-y3<α, the current new substrate defect class is added as a new child node, and the source node network is horizontally expanded, where β and α are thresholds selected by the learning algorithm.

7. The substrate defect detection method based on incremental learning CNN tree model according to claim 6 is characterized in that: The depth D and width constraint W of the maximum child node of the tree network are specifically expressed as: Among them, D max Indicates the preset maximum depth of a positive integer N + |children i | represents the number of child nodes on the same layer under node i in the tree network; And / or, setting the inductive preference of the tree network nodes based on the Occam's razor principle.

8. The substrate defect detection method based on incremental learning CNN tree model according to claim 1 is characterized in that: Based on the back propagation strategy at different network layer depths, the hidden layers of the substrate defect classification CNN network are retrained or fine-tuned using the newly added substrate defects and initial defect datasets, including: Add the substrate defect class to be added to the sub-node of the currently working sub-node CNN network, and reorganize the newly added substrate defects and initial training data; According to the three-dimensional likelihood matrix information of the newly added nodes, the hidden layer of the substrate defect classification CNN network is retrained or fine-tuned using a back propagation strategy based on different network depths; After the local CNN network parameters in the tree network are updated, the tree network enters the next batch of substrate defect classification and local CNN network model iterative update until no new substrate defect types are detected. The CNN network of each node of the tree network stops iterative update, and finally the substrate defect classification CNN network on each node of the trained tree network is obtained.

9. The substrate defect detection method based on incremental learning CNN tree model according to claim 8 is characterized in that: The back propagation strategy is expressed as: Among them, θ represents the parameters of the network, η is the learning rate, J represents the loss function, and x (i：i+n) Represents the characteristic values ​​of the i-th to i+n-th training samples, y (i：i+n) denotes the labels of the training samples of corresponding sizes, Represents the gradient of solving the network parameter θ.

10. The substrate defect detection method based on incremental learning CNN tree model according to claim 1, characterized in that: In the process of retraining or fine-tuning the hidden layer of the substrate defect classification CNN network, the Dropout layer is introduced to enhance regularization, which is specifically expressed as: Among them, II represents the indicator function, r i is a random variable that follows Bernoulli distribution Bernoulli(p), x i is the input of the i-th neuron, y i is the output of the i-th neuron after Dropout processing.

Images