A strip steel surface defect detection method and device based on deep learning

Abstract

The application discloses a kind of based on deep learning strip steel surface defect detection method and device, method includes: in strip steel production pipeline, the original image of strip steel surface is collected;The original image collected is marked with defect category, and the original image and marked data are integrated into strip steel surface defect dataset;A certain proportion of strip steel surface defect dataset is divided, and training set, verification set and test set are obtained;Strip steel surface defect detection model is constructed, and the model includes main network, focusing type feature pyramid network and detection head;The network is trained using strip steel surface defect dataset to obtain trained model, and the weight file in training process is saved;The trained model is used to detect strip steel surface defect image, and the type and position information of defect in image are obtained.The application effectively improves the precision and speed of strip steel surface defect detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of machine vision image processing, and in particular to a strip steel surface defect detection method and device based on deep learning. BACKGROUND

[0002] As one of the important raw materials of manufacturing industry, strip steel is widely used in the fields of automobile industry, aerospace and instrument manufacturing. At present, with the vigorous development of high-tech industries, the requirements of manufacturing industry for strip steel product quality are continuously improving. However, due to factors such as production equipment, production environment and manufacturing level, various defects will inevitably occur on the surface of strip steel during the production process. Using low-quality strip steel with defects for subsequent production and processing will easily cause hidden troubles, and even directly lead to scrap, which will damage the production efficiency of the overall industrial chain. Therefore, timely detection and processing of defects on the surface of strip steel can effectively control the defects of strip steel, reduce the economic losses of enterprises and promote the development of related industries. Therefore, the research on defect automatic detection algorithm based on deep learning has very important practical value for improving the quality of strip steel. In the prior art, the defect detection technology based on deep learning solves the problem of manual quality inspection to some extent, but in actual industrial scenarios, it is difficult to meet the actual demand, and the detection level is still insufficient: on the one hand, complex strip steel surface defects exist complex noise, and the algorithm is prone to missing detection and missed detection of defects. On the other hand, most of the existing algorithms are slow in overall detection speed in actual industrial scenarios with high real-time demand, which slows down the production process of strip steel. SUMMARY

[0003] In view of the problems of the prior art, the purpose of the present application is to provide a strip steel surface defect detection method and device based on deep learning, which effectively improves the precision and speed of strip steel surface defect detection.

[0004] The present application adopts the following technical solutions:

[0005] On the one hand, a strip steel surface defect detection method based on deep learning comprises:

[0006] Step S1, collecting original images of the surface of strip steel on a strip steel production line;

[0007] Step S2, labeling the collected original images according to defect categories, and integrating the original images and the labeled data into a strip steel surface defect data set;

[0008] Step S3, dividing the strip steel surface defect data set by a certain proportion to obtain a training set, a verification set and a test set;

[0009] Step S4, constructing a strip steel surface defect detection model, the model comprising a backbone network, a focal type feature pyramid network and a detection head;

[0010] Step S5, using the strip steel surface defect dataset to train the network to obtain a trained model, and saving the weight file in the training process;

[0011] Step S6, using the trained model to detect the strip steel surface defect image to obtain the type and location information of the defects in the image;

[0012] Among them, the backbone network extracts three different scale feature layers C1, C2 and C3, and transmits C1, C2 and C3 to the focal feature pyramid network; the focal feature pyramid network is used to complete the recalibration of the features in the neck network, and obtain feature layers P1, P2 and P3; in the detection head part, the feature layers P1, P2 and P3 are respectively cut to realize decoupled prediction.

[0013] Preferably, the collected original image is labeled for defect category, and the original image and the labeled data are integrated into a strip steel surface defect dataset, specifically including:

[0014] The labelImg software is used for data labeling to obtain a labeled file in XML format; the original image and the labeled data are integrated into a strip steel surface defect dataset.

[0015] Preferably, the backbone network includes image blocking, linear embedding, SWTR structure block and merging block, and four-stage feature extraction is performed; in the first stage, the input image is blocked and each image block is linearly embedded with a label using the image blocking and linear embedding, and the linear image block is extracted for features using the attention mechanism of the SWTR structure block; in the second stage, the transmitted feature layer is down-sampled using the merging block, and the feature layer is extracted for features using the SWTR structure block to obtain the feature layer C1; the operation of the second stage is repeated in the third and fourth stages to obtain the feature layers C2 and C3; the three different scale feature layers C1, C2 and C3 extracted by the backbone network are transmitted to the focal feature pyramid network.

[0016] Preferably, the focusing feature pyramid network comprises a CBS convolution module, an adaptive focusing block, an adaptive focusing residual block and an up-sampling; in the focusing feature pyramid network, three different scale feature layers are used for reinforced feature extraction; in terms of small feature layers of 7x7 pixels, two adaptive focusing blocks are used in succession on C3 and up-sampling is performed to obtain a feature layer C2'; one adaptive focusing block is used on C2' to perform up-sampling to obtain a feature layer P1'; one adaptive focusing block is used on C1 to perform channel adjustment, and one addition operation is performed with P1' to obtain a feature layer P1; in terms of medium feature layers of 14x14 pixels, one adaptive focusing block and one adaptive focusing residual block are used in succession on C2 to obtain a feature layer P2', and one addition operation is performed with C2' to obtain a feature layer P2; in terms of large feature layers of 28x28 pixels, one CBS convolution module is used on the adjusted C1 to perform down-sampling, and two adaptive focusing blocks are used in succession on the adjusted C1, then one adaptive focusing residual block is used on P2', a cascade operation is performed on the two, a channel adjustment of the CBS convolution module is performed on the feature layer, one adaptive focusing residual block is used again and an addition operation is performed with the adjusted feature layer of C3 to obtain a feature layer P3.

[0017] Preferably, in the detection head part, three YOLOXHead modules are used to respectively segment the feature layers P1, P2 and P3 to realize decoupled prediction, one side of the decoupled prediction uses a CBS convolution module with a 3x3 convolution kernel size and adjusts the channel to the number of defect types to predict the classification result; the other side uses a basic convolution module with a 3x3 convolution kernel size and respectively adjusts the channel to predict the positions of the four points of the anchor frame and determine the existence of the target.

[0018] Preferably, the loss function Loss used when training the strip surface defect detection model is:

[0019] Loss=λLoss reg +Loss cls +Loss obj

[0020] wherein, Loss reg represents a regression loss, which is calculated using an IoU loss; Loss cls represents a classification loss, Loss obj represents a target loss, all of which are calculated using a BCEWithLogits loss; λ is a weight value, and the optimizer uses Adam.

[0021] On the other hand, a strip surface defect detection device based on deep learning comprises:

[0022] An original image acquisition module is configured to acquire original images of a strip surface on a strip production line.

[0023] a defect dataset integration module for labeling the collected original images by defect categories and integrating the original images and the labeled data into a strip steel surface defect dataset;

[0024] a defect dataset division module for dividing the strip steel surface defect dataset by a certain proportion to obtain a training set, a verification set and a test set;

[0025] a model construction module for constructing a strip steel surface defect detection model, the model comprising a backbone network, a focal feature pyramid network and a detection head;

[0026] a model training module for training the network using the strip steel surface defect dataset to obtain a trained model and saving the weight file in the training process;

[0027] a strip steel surface detection module for detecting a strip steel surface defect image using the trained model to obtain the type and position information of the defect in the image.

[0028] Compared with the prior art, the present application has the following advantages:

[0029] 1. The present application uses an anchor-free YOLOX network and introduces SWTR as a backbone network, which accelerates the inference time in detection and fully improves the global modeling ability of the network, so that the network can more effectively capture the long-distance dependence of features, and finally realize more complete detection of the defect target area. This efficient combination can fully improve the detection efficiency of the detection equipment in the actual strip steel generation scene.

[0030] 2. The focal feature pyramid network designed in the present application can effectively improve the focusing of the network on the different scale features captured by SWTR. The overall pyramid structure based on convolution strengthens the flexibility of the network in feature capture, and at the same time, the defect features are locally connected and recalibrated to continuously adjust the focus of the network on the defects. The continuous pyramid transmission structure effectively fuses the characteristics possessed by the feature layers at different stages, significantly improves the local spatial aggregation ability of the detection network, and enables the network to accurately detect the position and type of the defect when dealing with strip steel surface defect targets with various shapes.

[0031] 3. The adaptive focusing block designed in the present application realizes the self-adaptation of the network to the shape of the strip steel surface defect and the recalibration of the features, making the network more flexible when facing defect targets and fully focusing on the effective features of the strip steel surface defect, solving the problem of missed detection and false detection when the network detects strip steel defect targets with large scale changes and much irrelevant noise. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1A flow chart of a strip steel surface defect detection method based on deep learning according to an embodiment of the present application;

[0033] Figure 2 A schematic diagram of a CBS convolution module according to an embodiment of the present application;

[0034] Figure 3 A schematic diagram of an adaptive feature extractor according to an embodiment of the present application;

[0035] Figure 4 A schematic diagram of an adaptive focus block according to an embodiment of the present application;

[0036] Figure 5 A schematic diagram of an adaptive focus residual block according to an embodiment of the present application;

[0037] Figure 6 A schematic diagram of a YOLOXHead module according to an embodiment of the present application;

[0038] Figure 7 A schematic diagram of an extrusion excitation block according to an embodiment of the present application;

[0039] Figure 8 A schematic diagram of a strip steel surface defect detection network according to an embodiment of the present application;

[0040] Figure 9 A structural block diagram of a strip steel surface defect detection device based on deep learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0041] The present application will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present application and not used to limit the scope of the present application. Furthermore, it should be understood that those skilled in the art can make various modifications or changes to the present application after reading the content taught by the present application, and these equivalent forms also fall within the scope of the appended claims of the present application.

[0042] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the present application will be described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0043] In the following description, many specific details are set forth in order to fully understand the present application. The described embodiments are merely some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description of the application herein is for describing particular embodiments only and is not intended to be limiting of the application.

[0045] The terms "first", "second", and "third" and the like, as used in the description of the application and the claims herein, are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. Except where the context clearly requires otherwise, terminology from the field of patent law and the IEEE standards dictionary is used.

[0046] Referring to Figure 1 The embodiment of the strip surface defect detection method based on deep learning shown includes the following steps:

[0047] Step S1, collecting original images of the strip surface on the strip production line.

[0048] Step S2, labeling the collected original images by defect category, and integrating the original images and the labeled data into a strip surface defect dataset.

[0049] Specifically, labelImg software is used for data labeling to obtain a labeled file in XML format. The original images and the labeled data are integrated into a strip surface defect dataset.

[0050] Step S3, dividing the strip surface defect dataset by a certain proportion to obtain a training set, a validation set, and a test set.

[0051] Step S4, constructing a strip surface defect detection model, which includes a backbone network, a focus-type feature pyramid network, and a detection head.

[0052] Specifically, the following basic modules are included:

[0053] The CBS convolution module includes convolution, batch normalization, and a Silu activation function, as shown in Figure 2 .

[0054] The adaptive feature extractor includes a CBS convolution module, a deformable convolution, and a squeeze-and-excitation block, as shown in Figure 3 .

[0055] The adaptive focus block includes a CBS convolution module and an adaptive feature extractor, as shown in Figure 4 .

[0056] Adaptive focus residual block, including adaptive focus block and one residual edge, see Figure 5 as shown.

[0057] Focusing feature pyramid network, including CBS convolution module, adaptive focus block, adaptive focus residual block and up-sampling, see Figure 6 as shown.

[0058] YOLOXHead module, including CBS convolution module, for decoupling classification and regression tasks and adjusting channels to realize prediction of defect type, target and position, see Figure 7 as shown.

[0059] Wherein, the backbone network SWTR (Swin Transformer) includes image blocking, linear embedding, SWTR structure block and merging block, for global self-attention on the image to realize efficient feature extraction and improve the global modeling ability of the overall network. The CBS convolution module includes convolution, batch normalization and activation function. For channel adjustment and preliminary feature extraction. The adaptive feature extractor includes a CBS convolution module, a deformable convolution and a squeeze-and-excitation block, for effectively focusing on irregular and widely distributed strip surface defect features, and realizing more comprehensive and complete capture of defects. The adaptive focus block includes a CBS convolution module and an adaptive feature extractor, for deep feature extraction of the network. The adaptive focus residual block includes an adaptive focus block and one residual edge, for stabilizing the expression of the output feature. The focusing feature pyramid network includes a CBS convolution module, an adaptive focus block, an adaptive focus residual block and an up-sampling, for establishing local connection on the feature layer to improve the network's capture of key defect features. The YOLOXHead module includes a CBS convolution module, for decoupling classification and regression tasks and adjusting channels to realize prediction of defect type, target and position.

[0060] See Figure 8 as shown, the constructed network model performs the following steps:

[0061] Step S4-1, three different scale feature layers are extracted by the backbone network Swin Transformer, specifically: the input image is divided into 4 blocks of 48-dimensional feature layers according to the input ratio. Then, linear embedding is performed on the 4 blocks of feature layers, the feature dimension of each block of feature layers is mapped to 96 dimensions, and full connection operation is performed. Then, the feature extraction is performed using the window-based attention mechanism in the SWTR structure block and the sliding window operation to obtain the initial feature layer. The initial feature layer is merged by blocks, and the feature layer is down-sampled and compressed across a single pixel to perform deep feature extraction. The stage block merging and SWTR structure block operation are repeated three times to obtain three different scale feature layers after feature extraction.

[0062] Step S4-2, focusing on the different scale features captured by SWTR through the neck network focused feature pyramid network, enhancing the expression of different scale features, specifically:

[0063] In terms of small feature layers of 7x7 pixels, two consecutive adaptive focusing blocks are used on C3 and up-sampling is performed to obtain feature layer C2'. C2' is continued to be used with 1 adaptive focusing block, and up-sampling is performed to obtain feature layer P1'. C1 is adjusted using 1 adaptive focusing block, and 1 addition operation is performed with P1' to obtain feature layer P1;

[0064] In terms of medium feature layers of 14x14 pixels, 1 adaptive focusing block and 1 adaptive focusing residual block are used on C2 to obtain feature layer P2', and 1 addition operation is performed between P2' and C2' to obtain feature layer P2;

[0065] In terms of large feature layers of 28x28 pixels, C1 is down-sampled by 1 CBS convolution module and is used twice in succession with adaptive focusing blocks. Then, 1 adaptive focusing residual block is used on P2', and the two are used in cascade operation. The channel of the feature layer is adjusted by the CBS convolution module, and 1 adaptive focusing residual block is used again and added with the adjusted feature layer of C3 to obtain feature layer P3. Thus, the recalibration of the feature in the neck network is completed.

[0066] Step S4-3, generate a strip steel surface defect detection model through YOLOXHead, specifically:

[0067] Three YOLOXHead modules are used to split feature layers P1, P2 and P3 respectively to realize decoupled prediction. One side uses a pair of 3x3 convolution kernel size CBS convolution modules and adjusts the channel to the number of defect types to predict the classification result; the other side uses a 3x3 convolution kernel size basic convolution module and adjusts the channel adaptively to predict the four point positions of the anchor box and determine the existence of the target.

[0068] Step S5, using the strip surface defect data set to train the network to obtain a trained model, save the weight file during the training process, specifically including:

[0069] Step S5-1, the strip surface defect data set obtained in step S3 is divided into a training set, a test set and a validation set in a ratio of 8:1:1;

[0070] Step S5-2, using the divided strip surface defect data set to train the network, generating a strip surface defect detection model;

[0071] The loss function Loss used when training the strip surface defect detection model is:

[0072] Loss=λLoss reg +Loss cls +Loss obj

[0073] Wherein, Loss reg represents the regression loss, which is calculated using the IoU loss; Loss cls represents the classification loss, Loss obj represents the target loss, all of which are calculated using the BCEWithLogits loss; λ is a weight, and the optimizer uses Adam.

[0074] The image input size is 224x224, the learning rate is 0.00005, the momentum is 0.9, the weight decay is 0.0005, and the batch size is 32.

[0075] Step S6, using the trained model to detect the strip surface defect image to obtain the type and position information of the defects in the image.

[0076] Referring to Figure 9 The embodiment also discloses a strip surface defect detection device based on deep learning, which comprises:

[0077] An original image acquisition module 901 is configured to acquire original images of a strip surface on a strip production line;

[0078] A defect data set integration module 902 is configured to label the acquired original images according to defect categories, and integrate the original images and the labeled data into a strip surface defect data set;

[0079] A defect data set division module 903 is configured to divide the strip surface defect data set in a certain proportion to obtain a training set, a validation set and a test set;

[0080] The model construction module 904 is configured to construct a strip steel surface defect detection model, the model comprising a backbone network, a focal feature pyramid network and a detection head.

[0081] The model training module 905 is configured to train the network using a strip steel surface defect dataset to obtain a trained model, and save a weight file in the training process.

[0082] The strip steel surface detection module 906 is configured to detect a strip steel surface defect image using the trained model to obtain type and position information of defects in the image.

[0083] It is apparent for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, but can be implemented in other concrete forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all aspects as illustrative and not restrictive, and the scope of the present application is defined by the appended claims rather than the above description, and it is intended to include all changes falling within the meaning and range of equivalents of the elements of the claims. Any reference signs in the claims should not be considered as limiting the involved claims. In addition, it is clear that the word "comprise" does not exclude other units or, the singular does not exclude the plural. The plurality of units or devices stated in the system claims can also be implemented by one unit or device through software or hardware.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application rather than limit the present application, and although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit of the technical solutions of the present application.

Claims

1. A method for detecting surface defects in strip steel based on deep learning, characterized in that, include: Step S1: Acquire raw images of the strip surface on the strip production line; Step S2: Label the defect categories of the acquired raw images and integrate the raw images and labeled data into a strip steel surface defect dataset. Step S3: Divide the strip surface defect dataset proportionally to obtain the training set, validation set, and test set; Step S4: Construct a surface defect detection model for strip steel. The model includes a backbone network, a focused feature pyramid network, and a detection head. Step S5: Use the strip steel surface defect dataset to train the network to obtain the trained model, and save the weight file during the training process. Step S6: Use the trained model to detect surface defects in the strip steel to obtain information on the type and location of defects in the image. The backbone network extracts three feature layers C1, C2, and C3 at different scales, and passes C1, C2, and C3 to the focused feature pyramid network. The focused feature pyramid network is used to complete the recalibration of features in the neck network to obtain feature layers P1, P2, and P3. In the detection head part, it is used to segment feature layers P1, P2, and P3 respectively to achieve decoupled prediction. The backbone network comprises image segmentation, linear embedding, SWTR structured blocks, and merging blocks, performing feature extraction in four stages. In the first stage, the input image is segmented using image segmentation and linear embedding, and a label is linearly embedded into each image block. The attention mechanism of the SWTR structured blocks is used to extract features from the linear image blocks. In the second stage, the merging blocks are used to downsample the passed feature layers, and the SWTR structured blocks are used to extract features from the feature layers, resulting in feature layer C1. The second stage is repeated in the third and fourth stages to obtain feature layers C2 and C3. The three feature layers C1, C2, and C3 extracted by the backbone network at different scales are then passed to a focused feature pyramid network. The focused feature pyramid network includes a CBS convolutional module, adaptive focusing blocks, adaptive focusing residual blocks, and upsampling. In the focused feature pyramid network, feature extraction is enhanced using three feature layers of different scales. For the small 7×7 pixel feature layer, C3 is subjected to two consecutive adaptive focusing blocks and upsampling to obtain feature layer C2'. C2' is then upsampled using one more adaptive focusing block to obtain feature layer P1'. C1 is then channel-adjusted using one adaptive focusing block and added to P1' to obtain feature layer P1. For the medium 14×14 pixel feature layer… For the feature layer P2, one consecutive adaptive focusing block and one adaptive focusing residual block are used on C2 to obtain feature layer P2'. P2' and C2' are then added together to obtain feature layer P2. For the large feature layer of 28×28 pixels, the adjusted C1 is downsampled by a CBS convolution module once and two consecutive adaptive focusing blocks are used on it. Then, one adaptive focusing residual block is used on P2'. The two are then concatenated to adjust the channels of the feature layer using the CBS convolution module. Another adaptive focusing residual block is used and added together with the feature layer adjusted by C3 to obtain feature layer P3.

2. The deep learning-based strip steel surface defect detection method according to claim 1, characterized in that, The process of labeling the acquired raw images with defect categories and integrating the raw images with the labeled data into a strip steel surface defect dataset specifically includes: The labelImg software was used to annotate the data, resulting in an XML file. The original images and the annotated data were then integrated into a strip surface defect dataset.

3. The deep learning-based strip steel surface defect detection method according to claim 1, characterized in that, In the detection head, three YOLOXHead modules are used to segment feature layers P1, P2, and P3 respectively to achieve decoupled prediction. On one side of the decoupling, a pair of CBS convolutional modules with a kernel size of 3×3 are used and the number of channels is adjusted to the number of defect types to predict the classification result; on the other side, a basic convolutional module with a kernel size of 3×3 is used and the channels are adaptively adjusted to predict the position of the four points of the anchor box and determine the existence of the target.

4. The deep learning-based strip surface defect detection method according to claim 1, characterized in that, The loss function Loss used during the training of the strip surface defect detection model is: = + + ； in, This represents the regression loss, calculated using the IoU loss. Represents classification loss, The target loss is represented by the BCEWithLogits loss, which is used for calculation. For weights, the optimizer uses Adam.

5. A deep learning-based strip steel surface defect detection device, characterized in that, include: The raw image acquisition module is used to acquire raw images of the strip surface on the strip production line; The defect dataset integration module is used to label the defect categories of the collected raw images and integrate the raw images and labeled data into a strip steel surface defect dataset. The defect dataset partitioning module is used to proportionally partition the strip steel surface defect dataset to obtain training set, validation set and test set; The model building module is used to build a surface defect detection model for strip steel. The model includes a backbone network, a focused feature pyramid network, and a detection head. The model training module is used to train the network using the strip steel surface defect dataset to obtain a trained model and save the weight file during the training process. The strip surface detection module is used to detect surface defects in strip images using a trained model, and to obtain information on the type and location of defects in the images. The backbone network includes image segmentation, linear embedding, SWTR structure blocks, and merging blocks, performing feature extraction in four stages. In the first stage, the input image is divided into blocks using image segmentation and linear embedding, and a label is linearly embedded into each image block. The attention mechanism of the SWTR structure block is used to extract features from the linear image blocks. In the second stage, the passed feature layer is downsampled using merging blocks, and the feature layer is extracted using the SWTR structure block to obtain feature layer C1. The operation of the second stage is repeated in the third and fourth stages to obtain feature layers C2 and C3. The three feature layers C1, C2 and C3 at different scales extracted by the backbone network are passed to the focused feature pyramid network. The focused feature pyramid network includes a CBS convolutional module, adaptive focusing blocks, adaptive focusing residual blocks, and upsampling. In the focused feature pyramid network, feature extraction is enhanced using three feature layers of different scales. For the small 7×7 pixel feature layer, C3 is subjected to two consecutive adaptive focusing blocks and upsampling to obtain feature layer C2'. C2' is then upsampled using one more adaptive focusing block to obtain feature layer P1'. C1 is then channel-adjusted using one adaptive focusing block and added to P1' to obtain feature layer P1. For the medium 14×14 pixel feature layer… For the feature layer P2, one consecutive adaptive focusing block and one adaptive focusing residual block are used on C2 to obtain feature layer P2'. P2' and C2' are then added together to obtain feature layer P2. For the large feature layer of 28×28 pixels, the adjusted C1 is downsampled by a CBS convolution module once and two consecutive adaptive focusing blocks are used on it. Then, one adaptive focusing residual block is used on P2'. The two are then concatenated to adjust the channels of the feature layer using the CBS convolution module. Another adaptive focusing residual block is used and added together with the feature layer adjusted by C3 to obtain feature layer P3.

Images