Silicon carbide growth defect detection method based on YOLOv7
By adding shallow output prediction head and attention mechanism to the YOLOv7 network, the problems of slow detection speed and low accuracy in the existing silicon carbide dislocation detection methods are solved, and more efficient and accurate dislocation detection is achieved, reducing errors Leak detection rate.
Patent Information
- Application Number
- CN202411218404.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2024-12-10
AI Technical Summary
The existing silicon carbide dislocation detection methods have problems such as slow detection speed, low accuracy, easy to be disturbed by the environment and frequent missed detection, especially for dislocation detection with large numbers of small targets.
The prediction head of shallow output is added to the YOLOv7 network and an attention mechanism is introduced. The improved network model is obtained through the training of the silicon carbide dislocation image dataset, which is used to identify and count three types of dislocation defects.
The accuracy of silicon carbide dislocation detection is improved, missed and missed detection conditions are reduced, detection speed and stability are improved, and detection capabilities for small targets are enhanced.
Smart Images
Figure CN119107310A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of target recognition technology, and in particular to a silicon carbide growth defect detection method based on YOLOv7. Background Art
[0002] Silicon carbide (SiC) is a third-generation semiconductor material with properties such as wide bandgap and high breakdown electric field. Due to its outstanding performance advantages, it has gradually been widely used in various industries and is suitable for the manufacture of high-temperature, high-frequency and high-power devices.
[0003] Since the atomic arrangement of crystalline materials is not strictly in accordance with the periodicity of the lattice during the production process, the silicon carbide materials actually produced will have defects of varying degrees. Among them, the existence of one-dimensional dislocation defects in silicon carbide seriously affects the performance of silicon carbide power devices, causing the degradation of device parameters, especially making it impossible to achieve the superior characteristics of silicon carbide high-power devices. Silicon carbide dislocation defects include screw dislocations, edge dislocations and basal plane dislocations. The morphology and size of dislocation defects vary greatly, and the number of dislocations is large and the density is high.
[0004] The current commonly used detection method is manual detection, but the number of silicon carbide dislocation defects is large, the manual detection speed is slow, and long-term manual counting will cause visual fatigue and reduce detection accuracy. The machine vision detection method using Halcon software or OpenCV template matching is easily affected by the environment, the detection results are unstable, and the detection speed needs to be improved.
[0005] After searching, the patent application number is 201210171066.8, the application date is May 29, 2012, and the invention name is: A method and system for detecting dislocations in crystalline silicon. This application adopts a high-definition imaging system to image the crystalline silicon to be detected, and obtains a crystalline silicon corrosion image, thereby obtaining the grayscale pixel ratio of the corrosion pits in the crystalline silicon corrosion image; according to the relationship between the grayscale pixel ratio of the corrosion pits and the dislocation density value, the dislocation density value of the crystalline silicon sample to be detected is obtained. The dislocation density of the crystalline silicon to be detected is calculated based on the data of the grayscale pixel ratio of the corrosion pits and the dislocation density value of the standard sample, which has problems such as low detection efficiency and easy false detection and missed detection.
[0006] For another example, the patent application number is 201811120149.8, the application date is September 25, 2018, and the invention name is: A method for detecting dislocations in silicon carbide. The application proposes to use a plasma torch with etching gas to scan the surface of the wafer, etch the surface of the wafer, and then use a microscope to observe the defects of the wafer after cooling. This method proposes a wafer surface etching method that is less harmful and easy to work on the assembly line, but does not improve the speed of dislocation image detection. Summary of the invention
[0007] 1. Technical problem to be solved by the invention
[0008] In view of the fact that the existing schemes for silicon carbide dislocation detection need to be improved, the present invention provides a method for silicon carbide growth defect detection based on YOLOv7. The present invention labels three dislocation defects in a silicon carbide dislocation image dataset, adds a shallow output prediction head to the YOLOv7 network and introduces an attention mechanism to obtain an improved network, trains the improved network with the silicon carbide dislocation image dataset, and realizes the recognition and counting of the three dislocations through the obtained optimal model.
[0009] 2. Technical solution
[0010] In order to achieve the above object, the technical solution provided by the present invention is:
[0011] A method for detecting silicon carbide growth defects based on YOLOv7 of the present invention comprises the following steps:
[0012] Step 1: annotating dislocation defects of silicon carbide dislocation images to create a silicon carbide dislocation image dataset;
[0013] Step 2: Add a prediction head from the shallow feature map to the YOLOv7 network;
[0014] Step 3: Introduce the attention mechanism in the YOLOv7 network after adding the prediction head from the shallow feature map;
[0015] Step 4: Use the silicon carbide dislocation image dataset to train the improved network to obtain the optimal network detection model;
[0016] Step 5: Detect the silicon carbide dislocation image through the obtained optimal model, identify and count the dislocations, and output the dislocation detection result map and the number of dislocations.
[0017] Furthermore, in step one, screw dislocations, edge dislocations and basal plane dislocations of the silicon carbide dislocation image are labeled by using Labelme software to produce a silicon carbide dislocation image data set.
[0018] Furthermore, in step 2, a shallow feature map is output from the first multi-branch stacking structure ELAN in the YOLOv7 network, and the shallow feature map is sequentially convolved, normalized and activated to obtain a new feature map; the shallow feature map output by the second multi-branch stacking structure ELAN in the YOLOv7 network is sequentially convolved, normalized, activated and up-sampled to make the feature map obtained in this layer have the same size as the new feature map output by the previous layer, and the two layers of feature maps are fused; the fused feature map is passed through the multi-branch stacking structure ELAN to obtain a feature map with more feature information, the feature map is down-sampled through a transition module MP containing two branches, and is fused again with the feature map of the next layer; and the fused feature map is assisted in training through the residual structure RepConv to obtain a new output feature map for image prediction.
[0019] Furthermore, in step three, the attention mechanism CBAM is introduced in the four feature layers output by the YOLOv7 backbone network. The introduced attention mechanism CBAM includes a channel attention mechanism and a spatial attention mechanism.
[0020] Furthermore, the channel attention mechanism in CBAM performs average pooling and maximum pooling on the input feature map respectively, processes the two pooled feature maps using a shared fully connected layer, and then performs an addition activation operation to obtain the final channel attention mechanism feature map.
[0021] Furthermore, the spatial attention mechanism in CBAM takes the features obtained by the channel attention mechanism as input, calculates the maximum and average values of the channel of each feature point, stacks the results into a 2×H×W channel, reduces the number of channels to 1 through convolution operation, and generates the spatial attention mechanism feature map through activation operation.
[0022] Furthermore, the spatial attention mechanism feature map is multiplied by the channel attention mechanism feature map as the output of the entire attention mechanism CBAM.
[0023] Furthermore, the activation function used by the attention mechanism CBAM is the Sigmoid function.
[0024] 3. Beneficial effects
[0025] Compared with the existing known technologies, the technical solution provided by the present invention has the following significant effects:
[0026] (1) The present invention provides a method for detecting silicon carbide growth defects based on YOLOv7, and proposes an improved YOLOv7 network model for silicon carbide dislocation detection. According to the characteristics of small and large numbers of edge dislocations and basal plane dislocations in silicon carbide dislocation images, a prediction head from a shallow feature map is added to the network, and the shallow feature map is fused with other layer feature maps. The shallow network is rich in information, which can improve the detection accuracy of small targets and reduce the number of false detections. This effectively improves the detection accuracy of silicon carbide dislocations and reduces missed detections and false detections.
[0027] (2) In the present invention, a method for detecting silicon carbide growth defects based on YOLOv7 is used. The image is extracted through the backbone network to obtain four output feature layers of different sizes. Different targets have different importance on each feature layer, and the neural network computing resources are limited. The four feature layers output by the backbone network introduce the attention mechanism CBAM, which can adaptively pay attention to the silicon carbide dislocation corrosion pits, reduce the interference of other impurities in the background on the detection, improve the network feature expression ability and the effective utilization of computing resources, and improve the accuracy of silicon carbide dislocation detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 It is a schematic diagram of the process of the present invention;
[0029] Figure 2 is a prediction head model diagram from a shallow feature map in the present invention;
[0030] Figure 3 This is the structural diagram of the CBAM attention mechanism of the present invention;
[0031] Figure 4 This is a silicon carbide dislocation detection network diagram of the present invention;
[0032] Figure 5 It is the test result diagram of the present invention;
[0033] Figure 6 Comparison of dislocation detection results of silicon carbide using different networks. DETAILED DESCRIPTION
[0034] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments.
[0035] Example 1
[0036] Combination Figure 1In this embodiment, a silicon carbide growth defect detection method based on YOLOv7 is provided. Three dislocation defects in a silicon carbide dislocation image dataset are annotated. A prediction head from a shallow feature map is added to the YOLOv7 network. An attention mechanism is introduced to obtain an improved network. The improved network is trained using the silicon carbide dislocation image dataset. The three dislocations are identified and counted using the obtained optimal model.
[0037] The specific steps are as follows:
[0038] Step 1: Use Labelme software to label the three dislocation defects (screw dislocation, edge dislocation and basal plane dislocation) of the silicon carbide dislocation image to create a silicon carbide dislocation image dataset.
[0039] Step 2: Add a prediction head from the shallow feature map to the YOLOv7 network;
[0040] Combination Figure 2 and Figure 4 , a shallow feature map is output from the first multi-branch stacking structure ELAN in the YOLOv7 network, and the shallow feature map is sequentially convolved, normalized and activated to obtain a new feature map. The shallow feature map output by the second multi-branch stacking structure ELAN in the YOLOv7 network is sequentially convolved, normalized, activated and upsampled to make the feature map obtained in this layer have the same size as the new feature map output in the previous layer, and the two layers of feature maps are fused. The fused feature map is passed through the multi-branch stacking structure ELAN to obtain a feature map with more feature information, and the feature map is downsampled through the transition module MP containing two branches and fused again with the feature map of the next layer. The fused feature map is assisted in training through the residual structure RepConv to obtain a new output feature map for image prediction.
[0041] Step 3: Introduce the attention mechanism into the network, which is a YOLOv7 network with four prediction heads after adding prediction heads from shallow feature maps.
[0042] Combination Figure 3 , the attention mechanism CBAM is introduced in the four feature layers output by the backbone network. The introduced attention mechanism CBAM (Convolutional Block Attention Module) includes channel attention mechanism and spatial attention mechanism.
[0043] The channel attention mechanism in CBAM performs average pooling and maximum pooling on the input feature map (H×W×C) respectively, uses a shared fully connected layer to process the two 1×1×C feature maps after pooling, and then performs an addition activation operation to obtain the final channel attention mechanism feature map.
[0044] The spatial attention mechanism in CBAM takes the features obtained by the channel attention mechanism as input, calculates the maximum and average values of the channel of each feature point, stacks the results into a 2×H×W channel, reduces the number of channels to 1 through convolution operation, and generates the spatial attention mechanism feature map through activation operation.
[0045] The attention mechanism is introduced into the network, and the activation function used is the Sigmoid function, which is expressed as:
[0046]
[0047] Multiply the spatial attention mechanism feature map with the channel attention mechanism feature map to obtain the output of the entire CBAM.
[0048] Step 4: Use the silicon carbide dislocation image dataset to train the improved network, divide the labeled dataset into a training set and a validation set in proportion, use the training set to train the improved network, and obtain the optimal network detection model.
[0049] Step 5: Combine Figure 5 and Figure 6 , the three dislocations are identified and counted through the obtained optimal model, the silicon carbide dislocation image is detected using the obtained optimal model, and the dislocation detection result map and the number of the three dislocations are output.
[0050] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.
Claims
1. A method for detecting silicon carbide growth defects based on YOLOv7, characterized in that: The steps are: Step 1: annotating dislocation defects of silicon carbide dislocation images to create a silicon carbide dislocation image dataset; Step 2: Add a prediction head from the shallow feature map to the YOLOv7 network; Step 3: Introduce the attention mechanism in the YOLOv7 network after adding the prediction head from the shallow feature map; Step 4: Use the silicon carbide dislocation image dataset to train the improved network to obtain the optimal network detection model; Step 5: Detect the silicon carbide dislocation image through the obtained optimal model, identify and count the dislocations, and output the dislocation detection result map and the number of dislocations.
2. The method for detecting silicon carbide growth defects based on YOLOv7 according to claim 1, characterized in that: Step 1: Label the screw dislocation, edge dislocation and basal plane dislocation of the silicon carbide dislocation image by using Labelme software to create a silicon carbide dislocation image data set.
3. A method for detecting silicon carbide growth defects based on YOLOv7 according to claim 1 or 2, characterized in that: Step 2: Output a shallow feature map from the first multi-branch stacking structure ELAN in the YOLOv7 network, perform convolution, normalization and activation operations on the shallow feature map in sequence to obtain a new feature map; perform convolution, normalization, activation and upsampling operations on the shallow feature map output by the second multi-branch stacking structure ELAN in the YOLOv7 network in sequence, so that the feature map obtained in this layer has the same size as the new feature map output in the previous layer, and fuse the two layers of feature maps; obtain a feature map with more feature information through the multi-branch stacking structure ELAN, downsample the feature map through a transition module MP containing two branches, and fuse it with the feature map of the next layer again; The fused feature map is then assisted in training through the residual structure RepConv to obtain a new output feature map for image prediction.
4. The method for detecting silicon carbide growth defects based on YOLOv7 according to claim 3, characterized in that: Step three introduces the attention mechanism CBAM into the four feature layers output by the YOLOv7 backbone network. The introduced attention mechanism CBAM includes channel attention mechanism and spatial attention mechanism.
5. The method for detecting silicon carbide growth defects based on YOLOv7 according to claim 4, characterized in that: The channel attention mechanism in CBAM performs average pooling and maximum pooling on the input feature map respectively, processes the two pooled feature maps using a shared fully connected layer, and then performs an addition activation operation to obtain the final channel attention mechanism feature map.
6. The method for detecting silicon carbide growth defects based on YOLOv7 according to claim 5, characterized in that: The spatial attention mechanism in CBAM takes the features obtained by the channel attention mechanism as input, calculates the maximum and average values of the channel of each feature point, stacks the results into a 2×H×W channel, reduces the number of channels to 1 through convolution operation, and generates the spatial attention mechanism feature map through activation operation.
7. The method for detecting silicon carbide growth defects based on YOLOv7 according to claim 6, characterized in that: The spatial attention mechanism feature map is multiplied by the channel attention mechanism feature map as the output of the entire attention mechanism CBAM.
8. The method for detecting silicon carbide growth defects based on YOLOv7 according to claim 7, characterized in that: The activation function used by the attention mechanism CBAM is the Sigmoid function.