YOLOv8 reflective battery shell surface defect detection method based on self-adaptive weight independent dual mechanisms

Abstract

The invention discloses a reflective battery shell surface defect detection method based on a self-adaptive preprocessing YOLOv8 model, and belongs to the field of battery shell defect detection. According to the method, the original data set on the surface of the shell is processed through bilateral filtering in combination with adaptive Gaussian filtering, adaptive threshold selection, morphological operation of'corrosion-expansion-corrosion 'and the like, image reflection noise is removed, the image quality is improved, and the identifiability of target defects is improved. An improved adaptive weight independent dual-mechanism global attention GAM module is integrated into a YOLOv8 feature fusion network, the attention weight is adaptively adjusted according to gray information and light and shade contrast of a shell image, a Mish activation function is used for replacing a SiLU function in an original model convolution module, and the degree of attention to target defects is improved. And the preprocessed image is input into an improved YOLOv8 detection model, so that the model training difficulty is reduced, and the recognition precision of the surface defects of the high-reflection shell is improved.

Description

Technical Field

[0001] The present invention relates to the field of surface defect detection of lithium iron phosphate battery shells, and in particular to a method for detecting the surface quality of a highly reflective battery shell. Technical Background

[0002] Lithium-ion batteries are a type of clean and renewable energy that can be reused. Among them, cylindrical lithium iron phosphate batteries are widely used due to their high safety performance, good high temperature performance, long cycle life, and environmental friendliness. The shell is an important component of cylindrical lithium batteries. Affected by the raw materials and processing technology, scratches, pits, water spots and other defects are prone to appear on the shell surface, which affects the battery performance and poses safety hazards such as explosions. Therefore, it is necessary to perform defect detection on the shell to remove unqualified products. At present, battery shell surface defect detection is mainly divided into two categories:

[0003] The first category is methods based on image processing. In 2022, Li Zhengchao of Shenzhen Zhixian Technology Co., Ltd. proposed a method for detecting surface defects of lithium battery casings (publication number: CN114757929A). The invention uses a template matching algorithm to eliminate part of the background of the image to be tested, and performs differential operations on the image processed by k-means clustering, edge detection, histogram equalization and other operations with the defect-free template to extract image features. This method makes full use of manual feature extraction methods to solve the problem of small sample training, but the image difference depends on the image quality, which may lead to misjudgment or missed detection.

[0004] In response to the above-mentioned problem that the use of image difference method may lead to misjudgment or missed detection, in 2023, Yang Zongying and others from Jiabaiyu (Nantong) Electronics Co., Ltd. proposed a battery casing defect recognition method (authorization number: CN116523923B). This invention determines the grayscale threshold corresponding to each defect through the bit layered map of the image to be tested, and segments the grayscale image to obtain a merged image containing the shadow area and light area corresponding to each defect. The type of each defect is identified according to the positional relationship between the shadow area and the light area corresponding to the same defect and the preset light source. This method avoids the situation where some defects are difficult to identify when the grayscale of the defective area changes less than that of the normal area, thereby improving the detection accuracy; however, some complex defects are still difficult to identify.

[0005] The second category is methods based on deep learning. In 2022, Hu Haibing and others from Hefei University of Technology proposed a real-time detection method for battery shell end face defects based on improved YOLOv5 (publication number: CN114972316A). This invention embeds the attention mechanism CBAM in the YOLOv5 backbone network, adopts a lightweight upsampling module CARAFE in the neck network, and replaces the feature pyramid network FPN-PAN structure with Bi-FPN. This method improves the accuracy of identifying complex defects of different types and with similar structures; but it is susceptible to noise and redundant features.

[0006] In response to the above problems, in 2023, Li Yunfeng of Beijing Miaoxiang Technology Co., Ltd. proposed a battery shell surface defect detection method and system based on deep learning (authorization announcement number: CN117110305B). The invention calculates the correlation value based on the Pearson correlation coefficient, mutual information and weighted summation method, and combines the stability value and correlation value for feature selection, which improves the accuracy and reliability of feature selection, but still does not solve the problem that the shell image is easily disturbed by reflective noise.

[0007] In view of the lack of literature on surface defect detection of highly reflective cylindrical battery shells, we searched for related patents similar to those on highly reflective battery shell detection. In 2018, Wang Peng and others from Tianjin University proposed a method for detecting high-reflective surface defects based on image processing and neural network classification (Announcement No.: CN108520274B). After background subtraction and denoising of the original image, this method preliminarily determines the parts that may have defects through feature extraction, and then inputs them into the neural network to determine whether they are true defects. This method reduces the probability of false detection in traditional image processing detection methods, but this method does not distinguish the defect type.

[0008] In response to the above problems, in 2023, Liu Jianhua and others from Changchun University of Science and Technology proposed a polarization image road target detection method based on improved YOLOv8 (Announcement No.: CN117351448B). This method constructs the first C2f module through the deformable convolution module, and the second C2f module through the re-parameter module. It improves YOLOv8 by combining the convolutional attention mechanism module CBAM and the loss function module Wise IoU. It can more accurately determine the category of reflective road surfaces, but this method is not suitable for cylindrical shells that are susceptible to uneven lighting.

[0009] According to the current research status, the research on battery shell surface defect detection methods has basically achieved real-time detection and defect location of common different types of battery shell surface defects; but the quality requirements for the collected images are high. When there is a large amount of high reflective noise on the shell surface and it is affected by uneven illumination, large detection errors are prone to occur. Therefore, the present invention proposes a reflective battery shell surface defect detection method based on the adaptive preprocessing YOLOv8 model. Bilateral filtering combined with adaptive Gaussian filtering is used to filter out part of the reflective noise while protecting the image edge details; through adaptive threshold selection, the image segmentation threshold is adaptively determined according to the pixel grayscale characteristics to avoid the arc shell surface being affected by uneven illumination. The improved GAM global attention mechanism is embedded in the YOLOv8 detection model to enhance the attention to small target defects on the shell surface and improve the detection accuracy. Summary of the invention

[0010] In order to overcome the shortcomings of existing technologies and methods, this paper proposes a reflective cell shell surface defect detection method based on the adaptive preprocessing YOLOv8 model. This method can realize defect detection of low-quality images of the surface of highly reflective cell shells collected in actual factory production.

[0011] The objective of the present invention is achieved through the following technical solutions, which are a method for detecting the surface quality of a highly reflective battery housing based on adaptive Gaussian filtering and threshold selection, and specifically include the following steps:

[0012] Step 1: Collect low-quality images of the highly reflective battery shell surface;

[0013] Step 2: Preprocess the collected original image of the battery shell: smooth the collected low-quality image through bilateral filtering combined with adaptive Gaussian filtering, filter out some reflection noise while protecting the image edge details; through adaptive threshold selection, adaptively determine the image segmentation threshold according to the pixel grayscale characteristics, and effectively extract the defect location; obtain the complete image contour through the "erosion-dilation-erosion" morphological operation.

[0014] Step 3: Build the YOLOv8 detection network: Use the MobileNetV3 network to replace the DarkNet53 in YOLOv8, and use its deep separable convolution and global average pooling to ensure the detection accuracy of the model; integrate the improved global attention mechanism GAM attention module into the feature fusion network to improve the network's attention to small target defects; replace the SiLU function in the original model convolution module with the Mish activation function to enhance the generalization ability of the model.

[0015] Step 4: Input the preprocessed image into the improved YOLOv8 model for defect detection.

[0016] The step 2 specifically includes: grayscale conversion of the collected low-quality image, and bilateral filtering of the converted grayscale image to remove large-area reflection noise while protecting image edge details.

[0017] Furthermore, adaptive Gaussian filtering is used for image preprocessing. The size of the convolution kernel window is determined according to the pixel grayscale information, and Gaussian blur processing is performed on the window area to remove the influence of noise;

[0018] Furthermore, the Sobel operator is used to calculate the gradient strength and direction of the filtered image; secondly, non-maximum suppression is used to compare the gradient value of the pixel with other points in the neighborhood. If the gradient strength of the current point is the largest, the gradient value is retained, otherwise the gradient of the point is 0;

[0019] Furthermore, the image segmentation threshold T is adaptively determined according to the grayscale characteristics of the image pixels. max and T min , if the pixel gradient is greater than T max , is an edge point; if the pixel gradient is less than T min , then discard it; if the pixel gradient is between the two, then find the pixel gradient value from the pixel 8 neighborhood, if there is a pixel gradient higher than T max , then it is an edge point. If there is none, it is discarded and the obtained edge is the defect position.

[0020] Furthermore, the image after threshold segmentation is subjected to a morphological operation of “erosion-dilation-erosion” to further remove noise and obtain a more complete image contour.

[0021] The step 3 specifically includes: building a YOLOv8 defect detection network. Adjusting and configuring the corresponding parameters to improve the detection results; using the MobileNetV3 network to replace the DarkNet53 in YOLOv8, and using its deep separable convolution and global average pooling to ensure the detection accuracy of the model; integrating the improved global attention mechanism GAM attention module into the feature fusion network to improve the network's attention to small target defects; replacing the SiLU function in the original model convolution module with the Mish activation function to enhance the model's generalization ability.

[0022] Furthermore, the execution process of replacing DarkNet53 in YOLOv8 with the MobileNetV3 network specifically includes:

[0023] The feature extraction network of the MobileNetV3 network is used to replace the 1st to 9th layers of the YOLOv8 backbone network. The deep separable convolution and global average pooling are used in the backbone network to ensure the detection accuracy of the model.

[0024] Furthermore, the improved global attention mechanism GAM attention module is integrated into the feature fusion network, specifically, at the connection between the backbone network and the neck network, the global attention mechanism GAM is added before the fully connected layers Concat_1, Concat_2, and Concat_3, so as to increase the attention to the target defects and improve the feature extraction and fusion capabilities of the network, so as to improve the detection accuracy of the target defects.

[0025] Furthermore, the improved global attention mechanism GAM specifically includes:

[0026] The feature map is input, and the number of channels is reduced through convolution with a convolution kernel of 5. Then the number of channels is increased through convolution with a convolution kernel of 5 to keep the number of channels consistent, and finally output through Sigmoid. A smaller convolution kernel is conducive to better capturing tiny features. The formula is as follows:

[0027] A S (F1)=σf 5×5 (f 5×5 (F1)) (1)

[0028] Among them, F1 is the input feature map, A S is the spatial attention weight, f 5×5 is a convolution kernel of size 5×5, and σ is batch normalization.

[0029] The channel attention mechanism and the spatial attention mechanism are "connected in parallel", that is, the feature map F1 is input into the channel attention mechanism and the spatial attention mechanism for separate processing, to avoid the mutual interference between the channel attention weight and the spatial attention weight, to avoid the loss of small target information, and to improve the network detection accuracy. The formula is as follows:

[0030]

[0031] Among them, F1 is the input feature map, A C is the channel attention weight, A S is the spatial attention weight, Represents the multiplication of corresponding elements of the matrix, and F2 is the output feature map.

[0032] In order to more reasonably adjust the distribution ratio of channel attention and spatial attention, the weight distribution coefficient α is introduced to redistribute according to the average grayscale value and light-dark contrast of the image. The greater the contrast of the battery shell image, the greater the brightness change of the image and the more obvious the details; in this case, the spatial attention can better focus on the key local detail features in the shell image. The formula is as follows:

[0033] α=tanh[k(I max -I avg)]gC(m,n) (3)

[0034] Among them, α is the attention mechanism weight allocation coefficient, I avg is the average gray value of the image, C(m,n) is the contrast of the image, k, I c All are constant terms.

[0035] The main problems in the detection of surface defects of highly reflective battery shells are as follows: a large amount of reflective noise is likely to exist on the surface of the metal shell, which affects the detection accuracy; the shell is cylindrical, the surface is unevenly illuminated, and the quality of the collected image is low; the defects on the shell surface vary in size, and small defects are easily missed. Therefore, the beneficial effects of the present invention are:

[0036] First, in order to address the problem that a large amount of reflective noise is easily present on the surface of metal shells, the present invention smoothes the collected original image through bilateral filtering combined with adaptive Gaussian filtering, thereby filtering out part of the reflective noise while protecting the image edge details; and performs a "corrosion-dilation-corrosion" morphological operation on the image after threshold segmentation to further remove noise while obtaining a complete image contour.

[0037] Second, in view of the uneven illumination on the surface of the cylindrical shell and the different reflection intensities that lead to different pixel values ​​at different positions of the shell, adaptive threshold selection is used to adaptively determine the image segmentation threshold according to the pixel grayscale characteristics to effectively extract the defect position.

[0038] Third, to address the problem that small defects on the shell surface are easily missed, the present invention integrates the improved adaptive weight independent dual-mechanism global attention GAM module into the feature fusion network of YOLOv8, and adaptively adjusts the attention weight according to the grayscale information and light and dark contrast of the shell image, thereby improving the network's attention to small defects on the surface of the battery shell; and replaces the SiLU function in the convolution module of the original model with the Mish activation function to enhance the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0040] Figure 2 This is a schematic diagram of the improved YOLOv8 network structure of the present invention;

[0041] Figure 3 Schematic diagram of the improved GAM global attention mechanism module of the present invention.

[0042] Figure 4 This is a schematic diagram of the improved convolution module of the present invention; DETAILED DESCRIPTION

[0043] The present invention is further described below in conjunction with the accompanying drawings and specific implementation methods:

[0044] like Figure 1 Shown is the overall flow chart of the present invention, comprising the following steps:

[0045] Step 1: Collect low-quality images with reflective noise on the surface of the reflective battery housing.

[0046] Step 2: Use image processing methods to preprocess the collected low-quality images and divide the data set;

[0047] Step 2-1: Convert the low-quality image collected in step 1 into a grayscale image and scale the image to an appropriate size to reduce the calculation time; use bilateral filtering to further remove the large area of ​​reflective noise on the shell surface and retain the shell edge details.

[0048] Step 2-2: Use adaptive Gaussian filtering to preprocess the image, determine the size of the convolution kernel window according to the pixel grayscale information, and perform Gaussian blur processing on the window area to remove the influence of noise;

[0049] Step 2-3: Use the Sobel operator to calculate the gradient strength and direction of the filtered image; use non-maximum suppression to compare the gradient value of the pixel with other points in the neighborhood to eliminate the points of the pseudo edge. If the gradient strength of the current point is the largest, then retain the gradient value, otherwise the gradient of the point is 0;

[0050] Step 2-4: Finally, adaptive threshold selection is used. The arc shell surface is unevenly illuminated and the reflection intensity is different, which leads to different pixel values ​​at different positions of the shell. The average grayscale value in the neighborhood centered on the current pixel is calculated according to the filter in step 2-2 as the reference threshold of the pixels in the area, and the proportional coefficient is set to obtain the segmentation threshold T min and T min , if the pixel gradient is greater than T max , is an edge point; if the pixel gradient is less than T min , then discard it; if the pixel gradient is between the two, then find the pixel gradient value from the pixel 8 neighborhood, if there is a pixel gradient higher than T max , then it is an edge point. If there is none, it is discarded and the obtained edge is the defect position.

[0051] Step 2-5: Perform the morphological operation of "erosion-dilation-erosion" on the image after adaptive threshold segmentation to further remove noise and obtain a more complete image contour. The original image after segmentation is recorded as A, and the structural element B is introduced:

[0052] Step 3: Build an improved YOLOv8 defect detection network (network structure as shown Figure 2 shown).

[0053] Step 3-1: Use MobileNetV3 network to replace DarkNet53 in YOLOv8, and use deep separable convolution and global average pooling to ensure the detection accuracy of the model; specifically, use the feature extraction network of MobileNetV3 network to replace the 1st to 9th layers of YOLOv8 backbone network, and use deep separable convolution and global average pooling in the backbone network to reduce the number of model parameters while ensuring the detection accuracy of the model;

[0054] Step 3-2: Integrate the global attention mechanism GAM attention module into the feature fusion network. Specifically, add the global attention mechanism GAM before the fully connected layers Concat_1, Concat_2, and Concat_3 at the connection between the backbone network and the neck network to increase the attention to the target defects and improve the network's feature extraction and fusion capabilities, so as to improve the detection accuracy of the target defects.

[0055] The improved global attention mechanism GAM specifically includes: inputting the feature map, reducing the number of channels through convolution with a convolution kernel of 5, increasing the number of channels through convolution with a convolution kernel of 5 to keep the number of channels consistent, and finally outputting through Sigmoid; a smaller convolution kernel is conducive to reducing noise and better capturing tiny features. The formula is as follows:

[0056] A S (F1)=σf 5×5 (f 5×5 (F1)) (4)

[0057] Among them, F1 is the input feature map, A S is the spatial attention weight, f 5×5 is a convolution kernel of size 5×5, and σ is batch normalization.

[0058] The channel attention mechanism and the spatial attention mechanism are "connected in parallel", that is, the feature map F1 is input into the channel attention mechanism and the spatial attention mechanism for separate processing, so as to avoid mutual interference between the channel attention weight and the spatial attention weight, avoid the loss of small target information, and improve the network detection accuracy.

[0059]

[0060] Among them, F1 is the input feature map, A C is the channel attention weight, A S is the spatial attention weight, Represents the multiplication of corresponding elements of the matrix, and F2 is the output feature map.

[0061] In order to more reasonably adjust the distribution ratio of channel attention and spatial attention, the weight distribution coefficient α is introduced to redistribute according to the average grayscale value and light-dark contrast of the image. The greater the contrast of the battery shell image, the greater the brightness change of the image and the more obvious the details; in this case, the spatial attention can better focus on the key local detail features in the shell image. The formula is as follows:

[0062] α=tanh[k(I max -I avg )]gC(m,n) (6)

[0063] Among them, α is the attention mechanism weight allocation coefficient, I avg is the average gray value of the image, I max is the maximum gray value of the image, C(m,n) is the contrast of the image, and k is a constant.

[0064] in,

[0065] Among them, C(m,n) is the image contrast, I max is the maximum gray value of the image, I min is the minimum grayscale value of the image.

[0066] Step 3-3: Use the Mish activation function to replace the SiLU function in the original convolution module to enhance the generalization ability of the model and improve the accuracy of small target defect detection.

[0067] Step 4: Input the image preprocessed in step 2 into the improved network model in step 3 for detection, and output a tested high-reflective battery shell surface quality detection model based on improved YOLOv8.

[0068] The present invention provides a reflective battery shell surface defect detection method based on the adaptive preprocessing YOLOv8 model. The collected original image is smoothed by bilateral filtering combined with adaptive Gaussian filtering to filter out part of the reflective noise while protecting the image edge details; the image after threshold segmentation is subjected to the morphological operation of "erosion-expansion-erosion" to further remove the noise while obtaining the complete image contour. Adaptive threshold selection is adopted to adaptively determine the image segmentation threshold according to the pixel grayscale characteristics, and effectively extract the defect position. By integrating the improved adaptive weight independent dual mechanism global attention GAM module into the feature fusion network of YOLOv8, the attention weight is adaptively adjusted according to the grayscale information and light and dark contrast of the shell image, so as to improve the network's attention to the small defects on the surface of the battery shell; and the SiLU function in the convolution module of the original model is replaced by the Mish activation function to enhance the generalization ability of the model.

Claims

1. A reflective cell shell surface defect detection method based on the adaptive preprocessing YOLOv8 model, characterized in that: The method inputs the preprocessed high-reflective shell surface image into the improved YOLOv8 model for defect detection, and the method includes the following steps: Step 1: A low-quality image with reflective noise on the surface of the reflective cell shell is collected by a polarization imaging method based on coaxial projection. Step 2: Use image processing methods to preprocess the collected low-quality images and divide the data set; Step 3: Build an improved YOLOv8 defect detection network. Step 4: Input the preprocessed image into the improved YOLOv8 model for defect detection.

2. The reflective cell shell surface defect detection method based on the adaptive preprocessing YOLOv8 model according to claim 1 is characterized in that: The specific steps in step 2 are as follows: 2-1 Convert the collected low-quality images into grayscale images and scale them to a suitable size to reduce the calculation time; 2-2 Use bilateral filtering combined with adaptive Gaussian filtering to smooth the image and filter out the reflection noise on the shell surface; 2-3 Use the Sobel operator to calculate the gradient strength and direction of the filtered image, and use non-maximum suppression to compare the gradient value of the pixel with other points in the neighborhood to eliminate the pseudo-edge points 2-4 Perform adaptive threshold segmentation on the image to avoid the impact of uneven lighting on the shell surface; 2-5 Perform the morphological operation of "erosion-dilation-erosion" on the image.

3. The reflective cell shell surface defect detection method based on the adaptive preprocessing YOLOv8 model according to claim 1 is characterized in that: The specific steps in step 3 are as follows: 3-1 Use MobileNetV3 network to replace the YOLOv8 backbone feature extraction network; 3-2 Introducing the GAM attention mechanism into the feature fusion network of YOLOv8; 3-3 Use the Mish activation function to replace the SiLU function in the original Conv convolution module.

4. According to claim 3, it is characterized in that: The detailed method of step 3-1 is as follows: The feature extraction network of the MobileNetV3 network is used to replace the 1st to 9th layers of the YOLOv8 backbone network. The depthwise separable convolution and global average pooling are used in the backbone network to reduce the number of model parameters while ensuring the detection accuracy of the model.

5. According to claim 3, it is characterized in that: The detailed method of step 3-2 is as follows: At the connection between the backbone network and the neck network, the global attention mechanism GAM is added before the fully connected layers Concat_1, Concat_2, and Concat_3 to increase the attention to the target defects and improve the feature extraction and fusion capabilities of the network, so as to improve the detection accuracy of the target defects. The improvements of the global attention mechanism GAM include: inputting the feature map, reducing the number of channels through convolution with a convolution kernel of 5, increasing the number of channels through convolution with a convolution kernel of 5 to keep the number of channels consistent, and finally outputting through Sigmoid; a smaller convolution kernel is conducive to reducing noise and better capturing tiny features. The formula is as follows: A S (F1)=σf 5×5 (f 5×5 (F1)) (1) Among them, F1 is the input feature map, A S is the spatial attention weight, f 5×5 is a convolution kernel of size 5×5, and σ is batch normalization. The channel attention mechanism and the spatial attention mechanism are "connected in parallel", that is, the feature map F1 is input into the channel attention mechanism and the spatial attention mechanism for separate processing, so as to avoid mutual interference between the channel attention weight and the spatial attention weight, avoid the loss of small target information, and improve the network detection accuracy. Among them, F1 is the input feature map, A C is the channel attention weight, A S is the spatial attention weight, Represents the multiplication of corresponding elements of the matrix, and F2 is the output feature map. In order to more reasonably adjust the distribution ratio of channel attention and spatial attention, the weight distribution coefficient α is introduced to redistribute according to the average grayscale value and light-dark contrast of the image. The greater the contrast of the battery shell image, the greater the brightness change of the image and the more obvious the details; in this case, the spatial attention can better focus on the key local detail features in the shell image. The formula is as follows: α=tanh[k(I max -I avg )]gC(m,n) (3) Among them, α is the weight distribution coefficient of the attention mechanism, I avg is the average gray value of the image, C(m,n) is the contrast of the image, k, I c All are constant terms.

Images