A method for detecting defects in an infrared thermal image of a battery insulating coating layer
By improving the YOLOv11 model and combining it with the RFABlock and RFAConv modules, the RFAMANet network was constructed, which solved the problems of insufficient accuracy and real-time performance in thermal imaging defect detection of battery insulation coatings, and achieved more efficient identification and detection of minute defects.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AUTOMOTIVE ENGINEERING CORPORATION
- Filing Date
- 2026-04-17
- Publication Date
- 2026-06-09
Smart Images
Figure CN122175960A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and industrial defect detection technology, and in particular to an infrared thermal imaging defect detection method for battery insulation coating layers. Background Technology
[0002] With the rapid development of the new energy vehicle industry, lithium-ion batteries have become the mainstream choice for automotive power batteries due to their high energy density and long cycle life. The battery insulation coating, as a key structure ensuring battery safety, directly affects the battery's insulation performance and operational safety. During production and use, the insulation coating is prone to defects such as dirt, scratches, and bulges. If these defects are not detected in time, they may lead to serious safety accidents such as short circuits and thermal runaway.
[0003] Currently, defect detection in the insulating coating of automotive power batteries mainly relies on manual visual inspection or traditional image processing techniques. However, these methods have significant shortcomings in terms of detection efficiency, consistency, and the ability to identify minute defects. In recent years, deep learning-based target detection algorithms (such as the YOLO series) have been widely used in industrial visual inspection. However, under thermal imaging conditions, defects in battery insulating coatings often exhibit characteristics such as small scale, diverse shapes, and low contrast with the background. Existing models are still insufficient in feature extraction, multi-scale fusion, and noise suppression, resulting in detection accuracy and real-time performance that cannot meet the needs of industrial sites. Summary of the Invention
[0004] To address the aforementioned problems, this invention provides an infrared thermal imaging defect detection method for battery insulation coating layers.
[0005] To achieve the above objectives, this application provides the following technical solution:
[0006] A method for detecting defects in battery insulation coating using infrared thermal imaging, characterized in that it includes:
[0007] Step 100: Use a thermal imaging acquisition device to acquire thermal imaging images of the insulating coating layer of the automotive power battery, and construct a balanced dataset containing normal samples and defective samples.
[0008] Step 200: Improve the original YOLOv11 model to obtain a new YOLOv11 model: First, replace the C3K2 module in the original YOLOv11 backbone network with the RFABlock module and the Conv module with the RFAConv module. Then, build the RFAMANet neck network by embedding the RFABlock module based on the MAFPN architecture, and replace the original YOLOv11 neck network.
[0009] Step 300: Perform end-to-end training on the RFAMANet neck network, using adaptive learning rate adjustment and early stopping strategies. After training, the improved YOLOv11 model is used for defect detection of the insulating coating layer of automotive power batteries under thermal imaging.
[0010] In step 100, the thermal imaging image acquisition steps include:
[0011] Step 101: Place the vehicle power battery to be tested directly below the infrared thermal imaging camera;
[0012] Step 102: Uniformly heat the surface of the vehicle's power battery;
[0013] Step 103: Take an infrared thermal imaging image of the insulating coating layer of the automotive power battery using an infrared thermal imaging camera to obtain the corresponding infrared thermal imaging image;
[0014] Step 104: Acquire and store the infrared thermal imaging image obtained in step 103 as the original input image for model training.
[0015] The RFABlock module includes a Conv module, an RFAConv module, and a Concat module. The Conv module includes a first convolutional layer and a second convolutional layer. The first convolutional layer of the Conv module is used to convolve the upstream feature map of the upstream network layer of the RFABlock module to obtain a preliminary processed feature map. The RFAConv module includes multiple cascaded RFAConv layers. Each layer optimizes the preliminary processed feature map through a receptive field attention mechanism to obtain an optimized feature map. The Concat module is a concatenation layer used to concatenate and fuse multiple optimized feature maps processed by the RFAConv layers.
[0016] The RFAConv module dynamically adjusts the parameters of the convolutional kernel within the receptive field using a receptive field attention mechanism. The specific steps include:
[0017] Step 201: Input the pre-processed feature map, then extract the receptive field spatial features through fast group convolution (GroupConv), and simultaneously perform global average pooling (AvgPool) on the original feature map to generate a new feature map. ;
[0018] Step 202: Process the feature map Perform 1×1 group convolution and Softmax function processing to generate receptive field attention maps. ;
[0019] Step 203: Transfer the attention map Characteristics of the sensory field space Multiplication yields the optimized feature map .
[0020] In step 201: Feature map ;
[0021] in, express Grouped convolutions of different sizes This indicates the initial processing of the feature map. This indicates a normalization operation.
[0022] In step 202: Attention map ;in, This represents a 1×1 grouped convolution.
[0023] The RFAMANet neck network includes two feature fusion branches: a bottom-up feature propagation branch and a top-down feature propagation branch. Both branches integrate SAF and AAF modules and use the RFABlock module to perform feature extraction. Each path includes two feature concatenation operations: the first concatenation corresponds to the SAF module, and the second concatenation corresponds to the AAF module.
[0024] The SAF module is a shallow-assisted fusion module used to fuse deep features with high-resolution shallow features while preserving localization details. The output of the SAF module is as follows: ;
[0025] in, This indicates the output of the SAF module; Indicates the SiLU activation function; This represents the number of control channels in a 1×1 convolution operation; This indicates a 3×3 downsampling convolution operation; Indicates an upsampling operation; This represents the backbone feature map output by the nth layer backbone network of the novel YOLOv11 after the S001 step improvement; This represents the backbone feature map output by the (n-1)th layer backbone network of the improved novel YOLOv11; This represents the backbone feature map output by the (n+1)th layer backbone network of the improved novel YOLOv11. Indicates the result after processing by the SAF module .
[0026] The AAF module is used to enhance the interactive utilization of multi-scale feature layer information, and the output of the AAF module is expressed as follows:
[0027] ;
[0028] in, This indicates the output of the AAF module; Indicates the result after processing by the SAF module ; This indicates the result after processing by the AAF module. .
[0029] The output feature maps of the SAF and AAF modules are both connected to the RFABlock module for feature enhancement processing. The mathematical expression for the feature enhancement operation is as follows:
[0030] ;
[0031] ;
[0032] in, This represents the primary attention enhancement feature, which is specifically achieved through... Obtained after applying RFABlock; This represents the second attention enhancement feature, which is specifically achieved through... Obtained after applying RFABlock.
[0033] Compared with the prior art, the beneficial technical effects of the present invention are as follows:
[0034] 1. In this invention, the embedded RFABlock and RFAConv adaptively enhance the feature response of key defects by means of the receptive field attention mechanism, while suppressing the interference of redundant information, thus ensuring accuracy while avoiding excessive consumption of computing resources.
[0035] 2. In RFAMANet, the MAFPN architecture is introduced. First, shallow detailed information is aggregated through shallow auxiliary fusion (SAF module), and then deep interaction of cross-layer features is achieved through advanced auxiliary fusion (AAF module), which improves the semantic richness of features.
[0036] 3. The SAF and AAF modules are connected to RFABlock to form a cascaded "fusion-enhancement" structure, which compensates for the insufficient capture of local context during multi-scale fusion, further optimizing the discriminative power of features and making the model more accurate in locating and classifying low-resolution defects. In summary, the model proposed in this paper not only achieves a comprehensive improvement in P, R, and mAP50:95 indicators in the thermal imaging battery insulation coating defect detection task, but also takes into account detection efficiency, making it more suitable for practical application needs. Attached Figure Description
[0037] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0038] Figure 1 This is a structural diagram of the new YOLOv11 model;
[0039] Figure 2 This is a structural diagram of the RFABlock module;
[0040] Figure 3 This is a comparison chart of YOLOv5, YOLOv8, YOLOv11 and the RFAMANet model;
[0041] Figure 4 This is a diagram showing the comparison of test results. Detailed Implementation
[0042] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] Example
[0044] This invention discloses a method for detecting defects in battery insulation coatings using infrared thermal imaging, comprising the following steps:
[0045] Step 1: Use a thermal imaging acquisition device to acquire thermal imaging images of the insulating coating layer of the automotive power battery, and construct a balanced dataset containing normal samples and defective samples.
[0046] The thermal imaging acquisition device uses an infrared thermal imaging camera, model MV-CI003-GL-T6 from Hikvision Robotics. The infrared thermal imaging camera is fixed to the bracket with bolts. The thermal imaging images cover normal sample images as well as sample images containing defects such as dirt, scratches, and bulges.
[0047] Specifically, the sample image, after data augmentation operations such as translation, rotation, and noise addition, contains a total of 1200 images, which are used for training, validation, and testing in an 8:1:1 ratio.
[0048] In this embodiment, the steps for acquiring thermal imaging images include:
[0049] Step 101: Place the vehicle power battery to be tested directly below the infrared thermal imaging camera;
[0050] Step 102: Use a hair dryer to uniformly heat the surface of the automotive power battery to highlight the differentiated characteristics of different types of defects under thermal imaging;
[0051] Step 103: Take an infrared thermal imaging image of the insulating coating layer of the automotive power battery using an infrared thermal imaging camera to obtain the corresponding infrared thermal imaging image;
[0052] Step 104: Use the MVS software provided by Hikvision to acquire and store the infrared thermal imaging image obtained in step 103, as the original input image for model training.
[0053] Step 2: Improve the original YOLOv11 model to obtain a new YOLOv11 model (also known as the RFAMANet model). This involves steps S001-S002, specifically:
[0054] S001: Replace the C3K2 module in the original YOLOv11 backbone network with the RFABlock module and the Conv module with the RFAConv module;
[0055] like Figure 2 As shown, the RFABlock module includes a Conv module, an RFAConv module, and a Concat module;
[0056] The Conv module includes a first convolutional layer and a second convolutional layer. The first convolutional layer of the Conv module is used to perform convolution operations on the upstream feature map of the upstream network layer of the RFABlock module to obtain a preliminary processed feature map.
[0057] The RFAConv module consists of multiple cascaded RFAConv layers. Each layer optimizes the initial feature map through a receptive field attention mechanism to obtain an optimized feature map.
[0058] The Concat module is a stitching layer used to stitch together and fuse multiple optimized feature maps processed by the RFAConv layer.
[0059] The second convolutional layer of the Conv module is used to integrate and output the feature maps fused by the Concat module;
[0060] The workflow is as follows: First, the original infrared thermal imaging input image is input into the first layer of the YOLOv11 network improved by step S001. After feature extraction by the upstream network layer of the RFABlock module, an upstream feature map is output. This upstream feature map is then fed into the first convolutional layer of the Conv module for convolution, and then sequentially passed through multiple RFAConv layers of the RFAConv module. Each RFAConv layer optimizes the feature map using a receptive field attention mechanism, and during this process, the network's ability to perceive and extract different local features is gradually enhanced. Then, the multiple feature maps processed by the multiple RFAConv layers are fused together through the concat module. Finally, the second convolutional layer of the Conv module integrates the fused feature maps, ultimately outputting an enhanced feature map. In this invention, RFAConv is introduced into the RFABlock module, combined with a spatial attention mechanism and receptive field spatial features. This module significantly enhances the model's adaptive extraction ability for defects at different scales and suppresses noise interference.
[0061] In this embodiment, the RFAConv module dynamically adjusts the convolutional kernel parameters within the receptive field using a receptive field attention mechanism. Specific steps include steps 201-203:
[0062] Step 201: Input the preliminary processed feature map, then extract the receptive field spatial features through fast group convolution (GroupConv), and simultaneously perform global average pooling (AvgPool) on the preliminary processed feature map to generate a new feature map. ;
[0063] Specifically, feature map ;
[0064] in, express Grouped convolutions of different sizes This indicates the initial processing of the feature map. Representing the normalization operation, RFAConv can dynamically generate different convolutional kernel weights for each receptive field, effectively solving the inherent limitation of parameter sharing in traditional convolution and enhancing the model's sensitivity to differences in local image location information. This method significantly improves the performance of convolutional neural networks in complex feature extraction tasks without increasing model parameters or computational overhead.
[0065] Step 202: Process the feature map Perform 1×1 group convolution and Softmax function processing to generate receptive field attention maps. Attention map ,in, This represents a grouped convolution of size 1×1;
[0066] Step 203: Transfer the attention map Characteristics of the sensory field space Multiplication yields the optimized feature map Among them, the optimized feature map ;
[0067] S002: Based on the MAFPN architecture, the RFABlock module is embedded within it to construct the RFAMANet neck network, replacing the original neck network in YOLOv11 with the RFAMANet neck network to achieve the following results: Figure 1 The new YOLOv11 model shown.
[0068] In step S001, the improved backbone network performs layer-by-layer feature extraction and downsampling on the input infrared thermal imaging image to generate four main network feature maps, P2, P4, P6 and P10, and inputs the above main network feature maps into the RFAMANet neck network.
[0069] Furthermore, the RFAMANet neck network includes two feature fusion branches: a bottom-up feature propagation branch and a top-down feature propagation branch.
[0070] The bottom-up feature propagation branch is the first feature propagation path of the RFAMANet neck network, and the top-down feature propagation branch is the second feature propagation path of the RFAMANet neck network. Both include the SAF module and the AAF module. The SAF module is a shallow auxiliary fusion module, which is used to fuse deep features with high-resolution shallow features, preserve localization details, and is responsible for extracting multi-scale features from the backbone network and completing the initial auxiliary feature fusion in the shallow neck layer. The AAF module is an advanced auxiliary fusion module, which is used to fuse feature information from multiple different levels, enhance the detection capability of medium-sized targets, collect gradient information from each layer through a denser connection structure, and finally guide the detection head to generate three different resolutions of diverse outputs.
[0071] In this embodiment, the core of the bottom-up feature propagation branch is the SAF module, and the core of the top-down feature propagation branch is the AAF module. Both the bottom-up and top-down feature propagation branches use the RFABlock module to perform feature extraction. This module can achieve adaptive adjustment of the receptive field by means of dynamically sized convolutional kernels. Each path includes two feature concatenation operations: the first concatenation corresponds to the SAF module, and the second concatenation is implemented by the AAF module.
[0072] The SAF module combines deep information with features at the same level and high-resolution shallow features to preserve rich localization details, thereby enhancing the network's spatial representation capabilities. Furthermore, the SAF module uses 1×1 convolutions to control the number of channels for shallow information, ensuring that shallow information occupies a small proportion during the concatenation process without affecting subsequent learning. Specifically, the output of the SAF module is as follows:
[0073] ;
[0074] in, This indicates the output of the SAF module; Indicates the SiLU activation function; This represents the number of control channels in a 1×1 convolution operation; This indicates a 3×3 downsampling convolution operation; Indicates an upsampling operation; This represents the backbone feature map output by the nth layer of the novel YOLOv11 backbone network after the S001 step improvement; specifically, it refers to a layer among the four multi-scale feature maps P2, P4, P6, and P10 ultimately output by this backbone network. Similarly, This represents the backbone feature map output by the (n-1)th layer of the improved YOLOv11 backbone network; similarly, This represents the backbone feature map output by the (n+1)th layer backbone network of the improved novel YOLOv11.
[0075] Indicates the result after processing by the SAF module That is, shallow high-resolution layers; Indicates the result after processing by the SAF module That is, a shallow, low-resolution layer.
[0076] The AAF module is used to enhance the interactive utilization of multi-scale feature layer information. Specifically, the output of the AAF module is expressed as follows:
[0077] ;
[0078] in, This indicates the output of the AAF module; This indicates the result after processing by the AAF module. ;
[0079] The AAF module integrates... , , , These four different levels of information achieve The AAF module aggregates information and uses 1×1 convolutions for channel control to adjust the influence weight of each layer on the output. Unlike the SAF module, the AAF module abandons the strategy of "half the number of channels in shallow layers" and unifies the number of channels in each layer. This design can preserve the initial guiding information contained in the shallow layers of MAFPN, thus ensuring that the model obtains rich output results.
[0080] Furthermore, in the RFAMANet neck network described in this invention, the output feature maps of the SAF module or AAF module are all connected to the RFABlock module for feature enhancement processing. Although the SAF module and AAF module have achieved effective integration of features at different levels, there may be problems such as insufficient representation of target details, background noise interference, and local feature alignment deviation in the fused features. However, the RFABlock module, through the receptive field attention mechanism of RFAConv, can dynamically adjust the convolution kernel parameters to adaptively enhance the feature response of key targets such as small defects, suppress redundant information, and refine the spatial structure and semantic expression of features. This makes up for the lack of local context capture in the multi-scale fusion process, and ultimately provides the detection head with more discriminative feature input, which meets the core requirements of detecting small defects in the thermal imaging battery insulation coating.
[0081] The above feature enhancement operation is as follows:
[0082] ;
[0083] ;
[0084] in, This represents the primary attention enhancement feature, which is specifically achieved through... Obtained after applying RFABlock; This represents the second attention enhancement feature, which is specifically achieved through... Obtained after applying RFABlock.
[0085] Step 3: Use the labeled thermal imaging images to train the RFAMANet neck network end-to-end, adopting adaptive learning rate adjustment and early stopping strategies. After training, the improved YOLOv11 model can detect defects in the insulating coating of automotive power batteries under thermal imaging.
[0086] In this embodiment, existing models YOLOv5, YOLOv8, and YOLOv11 are compared with the RFAMANet model. Precision (P), recall (R), and mAP50:95 are selected to evaluate model performance. mAP50:95 refers to the average precision of each class calculated at multiple thresholds with IoU ranging from 0.50 to 0.95 (in increments of 0.05), and the average of all average precisions is taken. The test results of different models on this self-built dataset are referenced. Figure 3 .
[0087] like Figure 3 Compared to existing models, the improved YOLOv11 model... , and The model achieved optimal performance across all metrics, demonstrating superior detection capabilities. This can be attributed to three main factors: First, to enhance the feature representation of low-resolution defects under thermal imaging, the embedded RFABlock and RFAConv adaptively enhance the feature response of key defects using a receptive field attention mechanism, while suppressing redundant information interference, ensuring accuracy while avoiding excessive consumption of computational resources. Second, to fully integrate multi-scale feature information, the MAFPN architecture is introduced in RFAMANet. Shallow auxiliary fusion (SAF module) aggregates shallow detail information, and then advanced auxiliary fusion (AAF module) achieves deep interaction of cross-layer features, improving the semantic richness of the features. Third, the SAF and AAF modules are followed by RFABlock to form a cascaded "fusion-enhancement" structure, compensating for insufficient local context capture during multi-scale fusion, further optimizing the discriminative power of the features, and making the model more accurate in locating and classifying low-resolution defects.
[0088] like Figure 4 As shown, the novel YOLOv11 model (also known as the RFAMANet model) obtained in this application can more completely identify various defects in the insulating coating compared to the original model. The defect detection is more comprehensive, the marking position of the defects is more accurate, and the overall detection effect is better.
[0089] In summary, the model proposed in this paper not only achieves the task of detecting defects in the thermal imaging battery insulation coating, but also... , and The comprehensive improvement of indicators, while taking into account detection efficiency, makes it more suitable for the application needs of real-world scenarios.
[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for detecting defects in battery insulation coating using infrared thermal imaging, characterized in that, include: Step 100: Acquire thermal imaging images of the insulating coating layer of automotive power batteries using a thermal imaging acquisition device, and construct a balanced dataset containing normal samples and defective samples; Step 200: Improve the original YOLOv11 model to obtain a new YOLOv11 model: first, replace the C3K2 module in the original YOLOv11 backbone network with the RFABlock module and the Conv module with the RFAConv module, and then construct the RFAMANet neck network by embedding the RFABlock module based on the MAFPN architecture, replacing the original YOLOv11 neck network; Step 300: Perform end-to-end training on the RFAMANet neck network, using adaptive learning rate adjustment and early stopping strategies. After training, the improved new YOLOv11 model is used for defect detection of the insulating coating layer of automotive power batteries under thermal imaging.
2. The method for detecting defects in battery insulation coating using infrared thermal imaging according to claim 1, characterized in that, In step 100, the thermal imaging image acquisition steps include: Step 101: placing the vehicle power battery to be tested directly below the infrared thermal imaging camera; Step 102: uniformly heating the surface of the vehicle power battery; Step 103: taking pictures of the insulating coating layer of the vehicle power battery with the infrared thermal imaging camera to obtain the corresponding infrared thermal imaging image; Step 104: acquiring and storing the infrared thermal imaging image obtained in step 103 as the original input image for model training.
3. The method for detecting defects in battery insulation coating using infrared thermal imaging according to claim 1, characterized in that, The RFABlock module includes a Conv module, an RFAConv module, and a Concat module. The Conv module includes a first convolutional layer and a second convolutional layer. The first convolutional layer of the Conv module is used to convolve the upstream feature map of the upstream network layer of the RFABlock module to obtain a preliminary processed feature map. The RFAConv module includes multiple cascaded RFAConv layers. Each layer optimizes the preliminary processed feature map through a receptive field attention mechanism to obtain an optimized feature map. The Concat module is a splicing layer used to splice and fuse multiple optimized feature maps processed by the RFAConv layer.
4. The method for detecting defects in battery insulation coating layer by infrared thermal imaging according to claim 3, characterized in that, The RFAConv module dynamically adjusts the parameters of the convolutional kernel within the receptive field using a receptive field attention mechanism. The specific steps include: Step 201: Input the pre-processed feature map, then extract the receptive field space features through fast group convolution, and simultaneously perform global average pooling on the original feature map to generate a new feature map. ; Step 202: Process the feature map Perform 1×1 group convolution and Softmax function processing to generate receptive field attention maps. ; Step 203: Transfer the attention map Characteristics of the sensory field space Multiplication yields the optimized feature map .
5. The method for detecting defects in a battery insulation coating using infrared thermal imaging according to claim 4, characterized in that, In step 201: Feature map ;in, express Grouped convolutions of different sizes This indicates the initial processing of the feature map. This indicates a normalization operation.
6. The method for detecting defects in battery insulation coating using infrared thermal imaging according to claim 4, characterized in that, In step 202: Attention map ;in, This represents a 1×1 grouped convolution.
7. The method for detecting defects in battery insulation coating layer using infrared thermal imaging according to claim 3, characterized in that, The RFAMANet neck network includes two feature fusion branches: a bottom-up feature propagation branch and a top-down feature propagation branch. Both branches integrate SAF and AAF modules and use the RFABlock module to perform feature extraction. Each path includes two feature concatenation operations: the first concatenation corresponds to the SAF module, and the second concatenation corresponds to the AAF module.
8. The method for detecting defects in battery insulation coating by infrared thermal imaging according to claim 7, characterized in that, The SAF module is a shallow-assisted fusion module used to fuse deep features with high-resolution shallow features while preserving localization details. The output of the SAF module is as follows: ;in, This indicates the output of the SAF module; Indicates the SiLU activation function; This represents the number of control channels in a 1×1 convolution operation; This indicates a 3×3 downsampling convolution operation; Indicates an upsampling operation; This represents the backbone feature map output by the nth layer backbone network of the novel YOLOv11 after the S001 step improvement; This represents the backbone feature map output by the (n-1)th layer backbone network of the improved novel YOLOv11; This represents the backbone feature map output by the (n+1)th layer backbone network of the improved novel YOLOv11. Indicates the result after processing by the SAF module .
9. The method for detecting defects in a battery insulation coating using infrared thermal imaging according to claim 7, characterized in that, The AAF module is used to enhance the interactive utilization of multi-scale feature layer information, and the output of the AAF module is expressed as follows: ;in, This indicates the output of the AAF module; Indicates the result after processing by the SAF module ; This indicates the result after processing by the AAF module. .
10. The method for detecting defects in a battery insulation coating using infrared thermal imaging according to claim 7, characterized in that, The output feature maps of the SAF and AAF modules are both connected to the RFABlock module for feature enhancement processing. The mathematical expression for the feature enhancement operation is as follows: ； ； in, This represents the primary attention enhancement feature, which is specifically achieved through... Obtained after applying RFABlock; This represents the second attention enhancement feature, which is specifically achieved through... Obtained after applying RFABlock.