Target detection method based on attention-driven feature enhancement

Abstract

The invention discloses a target detection method based on attention-driven feature enhancement, and relates to the field of computer vision, and the method comprises the steps: firstly inputting an image into a backbone network constructed based on YOLOv8; secondly, replacing one or more C2f modules in the backbone network with an MFEM module, carrying out feature enhancement and retention processing, and finally splicing the enhanced output features of a plurality of branches with the original features to obtain an enhanced feature map; then inputting the enhanced feature map into a JRAM module, performing channel and space attention weighting in parallel, retaining original features through residual connection, and outputting a weighted feature map; and finally, carrying out multi-scale feature fusion on the weighted feature map, inputting the fused features into a detection head, and outputting a target detection result. According to the method, the key problem of feature information loss in small target detection is effectively solved, fine features in the image can be better captured and reserved, and the recognition capability of a small-size target is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and more particularly to a target detection method based on attention-driven feature enhancement. Background Technology

[0002] Object detection is a key task in computer vision, aiming to identify the categories and locations of objects in images or videos. It has important significance in many practical applications, including autonomous driving, infrared detection, X-ray detection, pedestrian detection, robotics, medical imaging, and surveillance.

[0003] In recent years, deep learning-based object detection technologies (such as YOLO, SSD, Fast R-CNN, and Faster R-CNN) have made significant progress. However, small objects are extremely difficult to detect due to their low pixel count, inconspicuous features, and susceptibility to background interference. Compared to conventional object detection, small object detection faces challenges such as low resolution and insufficient semantic information, requiring more advanced algorithms and model architectures to extract subtle features. Traditional networks tend to lose subtle features of small objects during layer-by-layer downsampling; cross-stage partial connectivity (CSP) structures struggle to fully utilize all feature information.

[0004] Existing attention modules like CBAM rely on global pooling, leading to the loss of local details and high computational cost. Therefore, this invention proposes a novel attention mechanism inspired by CBAM to alleviate this problem, enabling the detection model to improve the processing of detailed information during layer-by-layer feature extraction, fully utilize as many feature information as possible in the image, and further refine the processing of small targets or fine-grained features. Summary of the Invention

[0005] The technical problem to be solved by this invention is to propose an Attention-Driven Feature Enhancement Network (AFENet) method and system, which can alleviate the defects in the processing of detailed information during the layer-by-layer feature extraction process, alleviate the problem of local detail loss in the processing of small targets or fine-grained features due to the reliance of spatial attention on global pooling, and alleviate the problem that cross-stage partial connections cannot fully utilize all feature information, thereby improving the performance of target detection.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a target detection method based on attention-driven feature enhancement, comprising the following steps:

[0007] S1. Input the preprocessed image into the backbone network built based on YOLOv8;

[0008] S2. Replace one or more C2f modules in the backbone network with a multi-branch feature preservation and enhancement (MFEM) module to perform feature enhancement and preservation processing. The MFEM module divides the input feature map into two parts: the first part directly retains the original features, and the second part is extracted through multi-branch feature enhancement. Finally, the output features enhanced by multiple branches are concatenated with the original features of the first part to obtain the enhanced feature map.

[0009] S3. Input the enhanced feature map into the Joint Residual Attention Mechanism (JRAM) module, perform channel and spatial attention weighting in parallel, retain the original features through residual connections, and output the weighted feature map.

[0010] S4. Input the weighted feature map into the feature pyramid network for multi-scale feature fusion, input the fused features into the detection head, and output the target detection result.

[0011] Further, in step S2, the processing procedure of the MFEM module includes:

[0012] The input feature map is divided into two parts: one part retains the original features, and the other part is enhanced and extracted through multiple branches. Each branch is used to extract and fuse features through the RFACA module and the PConv module in sequence. Finally, the output features of multiple branches are concatenated with the original features, and the number of channels is adjusted through a convolutional layer before output.

[0013] Furthermore, the specific processing flow of the MFEM module is as follows:

[0014] (1) First, adjust the number of channels of the input feature map to fit the dimension of subsequent processing by using 1×1 convolution;

[0015] (2) Divide the feature map after channel adjustment into branches C1 and C5. C5 directly retains the original features, and C1 is further split into parallel branches C2, C3 and C4.

[0016] (3) After the attention-weighted features of branches C2, C3 and C4 are extracted by the RFA module, the original features within the branches are retained by the PConv module and the enhanced features output by RFA are fused to obtain the enhanced features of the sub-branches.

[0017] (4) The output features of C2, C3, and C4 are concatenated with the original features of C1 and C5 through the Concat operation, and then fused by 1×1 convolution and the number of channels is adjusted to obtain the final enhanced features of the MFEM module.

[0018] Furthermore, the RFACA module is composed of RFAConv and a coordinate attention mechanism, which is used to provide attention weights for the convolution kernel and capture spatial features.

[0019] Further, in step S4, the processing procedure of the JRAM module includes:

[0020] The input feature maps are fed into the SEA and SA sub-modules for parallel processing. The SEA sub-module is used to calculate the channel dimension weights, and the SA sub-module is used to calculate the spatial dimension weights. Then, the outputs of the two sub-modules are residually concatenated with the original features of the input JRAM module. Finally, the two residuals are concatenated, and the output is adjusted by passing a convolutional layer.

[0021] Furthermore, in step S4, the JRAM module is deployed at layers 5, 8, and 11 of the backbone network.

[0022] Furthermore, the attention weight parameters of each JRAM module are independently optimized to adapt to the target feature enhancement requirements at the corresponding resolution.

[0023] The present invention also relates to an electronic device, the electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement an attention-driven feature enhancement-based target detection method of the present invention.

[0024] The present invention also relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an attention-driven feature enhancement-based target detection method of the present invention.

[0025] According to the above technical solution, the beneficial effects of the present invention are:

[0026] 1. In terms of detection performance, the key problem of feature information loss in small target detection is effectively solved through the synergistic effect of the Multi-Branch Feature Preservation Enhancement Module (MFEM) and the Joint Residual Attention Mechanism (JRAM). It can better capture and preserve subtle features in images, significantly improving the ability to recognize small targets and resulting in a marked improvement in the model's detection accuracy in complex scenes.

[0027] 2. In terms of computational efficiency, the innovative parallel attention architecture avoids the computational redundancy of traditional attention mechanisms, ensuring performance improvement while keeping computational resource consumption within a reasonable range. The modular design allows the network to be flexibly adjusted according to actual needs, achieving a balance between detection accuracy and operational efficiency.

[0028] 3. In terms of practical value, the improved network structure of this invention has strong generalization ability. It is not only suitable for conventional target detection tasks, but also performs well in small target detection scenarios in professional fields such as UAV aerial photography and medical imaging. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall architecture of the AFENet network of the present invention;

[0030] Figure 2 This is a schematic diagram of the MFEM module structure;

[0031] Figure 3 This is a schematic diagram of the RFACA module structure;

[0032] Figure 4 This is a schematic diagram of the JRAM module structure;

[0033] Figure 5 A schematic diagram illustrating the ablation experiment results of each module of AFENet on the VOC dataset;

[0034] Figure 6 This is a comparison of the qualitative results of AFENet, YOLOv8-N, and YOLOv12-N on the VisDrone2019 validation set. Detailed Implementation

[0035] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0036] The structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Without affecting the effects and objectives that the present invention can produce, they should still fall within the scope of the technical content disclosed in the present invention.

[0037] Combination Figure 1-4 As shown, this invention proposes a target detection method based on attention-driven feature enhancement, comprising the following steps:

[0038] S1. Input the preprocessed image into the backbone network built based on YOLOv8.

[0039] S2. Replace one or more C2f modules in the backbone network with a multi-branch feature preservation enhancement (MFEM) module to perform feature enhancement and preservation processing.

[0040] The specific processing flow of the MFEM module is as follows:

[0041] (1) First, adjust the number of channels of the input feature map to fit the dimension of subsequent processing by using 1×1 convolution;

[0042] (2) Divide the feature map after channel adjustment into branches C1 and C5. C5 directly retains the original features, and C1 is further split into parallel branches C2, C3 and C4.

[0043] (3) After the attention-weighted features of branches C2, C3 and C4 are extracted by the RFA module, the original features within the branches are retained by the PConv module and the enhanced features output by RFA are fused to obtain the enhanced features of the sub-branches.

[0044] Specifically, after C1 enters the RFACA module, it is divided into two parts. After being processed by average pooling and group convolution, the features of the two parts are reweighted and their shapes are adjusted. Then, they are processed by coordinate attention (CA) and finally, the final output parallel branches C2, C3, and C4 are obtained through convolution.

[0045] (4) The output features of C2, C3, and C4 are concatenated with the original features of C1 and C5 through the Concat operation, and then fused by 1×1 convolution and the number of channels is adjusted to obtain the final enhanced features of the MFEM module.

[0046] Specifically, the formula for calculating the output features of the MFEM module is as follows:

[0047] F(MFEM)=Conv(Concat(C1, C2, C3, C4, C5))

[0048] In this formula, F(MFEM) is the output feature of MFEM, where C1 and C5 represent two parts of the feature map obtained after convolution of the input x. C2, C3, and C4 represent the feature maps obtained after C1 is processed by RFACA and PConv, respectively, using different numbers of channels. Finally, these feature maps are concatenated using a Concat operation and then fused by convolution.

[0049] Furthermore, the RFACA module is composed of RFAConv and a coordinate attention mechanism, which is used to provide attention weights to the convolution kernel and capture spatial features.

[0050] S3. Input the enhanced feature map into the Joint Residual Attention Mechanism (JRAM) module, perform channel and spatial attention weighting in parallel, retain the original features through residual connections, and output the weighted feature map.

[0051] The processing steps of the JRAM module include:

[0052] The enhanced feature map from the MFEM module is input into the Joint Residual Attention (JRAM) module, and then further processed in parallel by the SEA and SA sub-modules. The SEA sub-module calculates the channel dimension weights, and the SA sub-module calculates the spatial dimension weights. Channel weight feature maps (C6) and spatial weight feature maps (C7) are then obtained. C6 and C7 are then residually concatenated with the original features input to the JRAM. Finally, the two residually concatenated feature maps are concatenated, and the number of channels is adjusted using a 1×1 convolution to output the semantically enhanced features.

[0053] Specifically, the formula for calculating the output characteristics of the JRAM module is as follows:

[0054] F(JRAM)=Conv(Concat(Add(x,C6),Add(x,C7)))

[0055] In the formula, F(JRAM) is the output feature of the JRAM module, where x represents the input, and C6 and C7 represent the two feature maps of the input x after being processed by the SEA and SA attention mechanisms, respectively. The feature maps of x are fused with C6 and C7 respectively, and finally these fused feature maps are concatenated by the Concat operation and then convolutional fusion operation is performed to obtain the final output of the JRAM module.

[0056] Preferably, JRAM modules are deployed at layers 5, 8, and 11 of the backbone network. The attention weight parameters of each JRAM module are independently optimized to adapt to the target feature enhancement requirements at the corresponding resolution.

[0057] S4. Input the weighted feature map into the feature pyramid network for multi-scale feature fusion, input the fused features into the detection head, and output the target detection result.

[0058] In the ablation, quantitative, and qualitative experiments conducted in this invention, a GeForce RTX 4090 GPU was used for both training and prediction. The environment framework used was PyTorch-2.2.2, and the CUDA environment was 11.8. Before training, the input image size was uniformly converted to 640×640. During training, no pre-trained weights were used, allowing the model to fit from scratch to fully utilize its potential performance. Furthermore, this invention also used the SGD optimizer for training, with an initial learning rate of 0.01 and a weight decay of 5×10⁻⁶. -4 Mosaic data augmentation was employed, with a batch size of 32 on the COCO2017 dataset, and training for 500 epochs. Various metrics were monitored during training, and the model that performed best on the validation set was selected.

[0059] The specific dataset and experiments are as follows:

[0060] (1) Datasets: COCO2017, PASCAL VOC07+12, VisDrone2019 datasets, covering 80, 20 and 10 target classes, with VisDrone2019 focusing on small target detection.

[0061] (2) Evaluation metrics: Evaluation criteria for the Microsoft COCO dataset: Average accuracy (AP), average accuracy for small targets (APS), average accuracy for medium-sized targets (APM), and average accuracy for large targets (APL).

[0062] (3) Ablation experiments: In order to verify the effectiveness of the modules of the present invention, the modules of AFENet were subjected to ablation experiments on the Microsoft COCO2017, PASCALVOC07+12 and VisDrone2019 datasets. The experimental results are shown in Tables 1-3. At the same time, the attention-driven model architecture based on the JRAM module was subjected to ablation experiments on the VOC07+12 dataset. The experimental results are shown in Table 4.

[0063] Table 1 Ablation experiments of each module of AFENet on the Microsoft COCO2017 dataset.

[0064]

[0065] Table 1 shows the experimental results based on the Microsoft COCO2017 dataset. After incorporating the evaluation metrics of the Microsoft COCO dataset, the baseline model YOLOv8n achieved an accuracy (AP50:95) of 37.4% on the Microsoft COCO2017 dataset. After adding the MFEM module, the accuracy (AP50:95) improved to 39.8%, and the accuracy (APS) for small targets improved by 1.7%. Specifically, the model architecture was used in the experiments. The accuracy (AP50:95) improved from 37.4% of the baseline model to 40.5%. The accuracy (APS) for small targets improved by 2.3%. After simultaneously implementing all modules of AFENet, the model accuracy (AP50:95) reached 43.8%, which is 6.4 percentage points higher than the baseline model's 37.4%. The accuracy (APS) for small targets improved by 5.7%, a significant improvement.

[0066] Table 2 Ablation experiments of each module of AFENet on the VOC dataset.

[0067]

[0068]

[0069] Continued from Table 2: Ablation experiments of each AFENet module on the VOC dataset.

[0070] method accuracy dining table dog horse motorcycle people Potted plants sheep sofa train television Baseline 79.8 77.1 83.5 89.1 88.1 87.9 54.3 77.1 75.6 88.4 79.7 +MFEM 81.7 79.2 85.4 91.8 88.9 88.6 56.6 83.0 78.4 88.6 80.7 +JRAM 82.5 79.1 87.6 91.6 89.0 89.0 57.9 81.9 78.3 88.6 84.2 AFENet 83.0 77.4 88.2 92.4 89.9 89.9 60.0 81.1 79.1 90.9 82.1

[0071] Table 2 shows that the baseline model YOLOv8n achieved an accuracy (mAP) of 79.8% on the PASCAL VOC07+12 dataset. After adding the MFEM module, the model's accuracy improved to 81.7%. For example, the accuracy (mAP) for small objects such as birds, dogs, potted plants, and sheep improved by 2.7%, 1.7%, 2.4%, and 5.9%, respectively. Building on the above, adding the JRAM module and introducing an attention-driven model architecture further improved the accuracy (mAP) from 79.8% to 82.5%, with a significant improvement in accuracy for small objects.

[0072] Overall, the accuracy of small object detection has been significantly improved after introducing the MFEM module, JRAM module and attention-driven model architecture into the baseline model YOLOv8n.

[0073] Table 3 Ablation experiments of each module of AFENet on the VisDrone2019 dataset.

[0074] method Number of parameters (M) Computational complexity (G) AP50 AP50:95 Ped. person bicycle Baseline 3.2 8.1 33.0 19.1 34.7 27.3 8.9 +MFEM 3.7 10.8 37.1 21.9 38.9 29.9 10.5 +JRAM 6.4 16.6 36.6 21.5 39.0 29.4 10.3 AFENet 7.2 18.7 38.2 22.7 40.8 31.2 10.6

[0075] Continued from Table 3: Ablation experiments of each AFENet module on the VisDrone2019 dataset.

[0076] method car van truck tricycle awn. bus motor Baseline 75.5 38.2 28.3 21.3 11.8 46.8 37.2 +MFEM 78.4 43.3 33.5 26.3 13.7 53.2 42.5 +JRAM 78.0 42.3 33.7 24.9 15.0 53.3 40.1 AFENet 78.6 43.8 36.7 26.3 15.4 53.6 42.6

[0077] Table 3 shows the parameters, computational cost, AP50, AP50:95, and accuracy (AP50) of AFENet on the VisDrone2019 dataset, as well as the accuracy (AP50) for all 10 categories. After adding the MFEM module, the parameters and computational cost slightly increased, but the accuracy (AP50) jumped from 33.0 to 37.1, an improvement of 4.1%. Building on this, adding the JRAM module further increased the parameters and computational cost, improving the accuracy (AP50) from 33.0 to 36.6, an improvement of 3.6%. The accuracy (AP50:95) increased from 19.1 to 21.5, an improvement of 2.4%. After introducing all the modules of AFENet, the accuracy (AP50) increased to 38.2, an improvement of 5.2%. The accuracy (AP50:95) increased to 22.7, an improvement of 3.6%. Furthermore, the highest accuracy (AP50) was achieved for all 10 categories. Although the number of parameters and computational load have increased, the improved accuracy verifies the effectiveness of this invention in small target detection.

[0078] Table 4 shows the ablation experiments of the JRAM-based attention-driven model architecture on the VOC07+12 dataset.

[0079] Position and number Number of parameters (M) Accuracy (%) Baseline 3.2 79.8 Backbone network, 1 4.6 79.7 Backbone network, 3 6.4 82.5 Neck, 1 4.6 80.9 Neck, 3 5.7 80.6

[0080] Research shows that directly using attention mechanisms in a model does not always achieve its intended effect. Therefore, to ensure the JRAM attention mechanism performs optimally in the model, Table 4 shows the parameter count and accuracy metrics of the JRAM-based attention-driven model architecture on the VOC07+12 dataset, enabling the attention mechanism to focus on feature weights at different resolutions. It can be seen that using the attention mechanism to process features at three resolution levels in the backbone network maximizes its performance; and the resulting performance improvement is significantly better than the other three methods using attention mechanisms.

[0081] (4) Comparative experiments: The present invention was compared with object detection methods on the PASCALVOC07+12 dataset, Microsoft COCO2017 dataset and VisDrone2019 dataset. The experimental results are shown in Table 5-7.

[0082] Table 5 compares AFENet with 11 other methods and benchmark methods on the VOC07+12 dataset.

[0083]

[0084]

[0085] Table 5 continues the comparison of AFENet with 11 other methods and benchmark methods on the VOC07+12 dataset.

[0086] Table 5 shows a comparison of the proposed AFENet with 11 other methods and benchmark methods on the PASCAL VOC07+12 dataset. The results represent the reimplementation of this invention; other data are from the original paper, and bold indicates the maximum value in that column. Table 5 shows that AFENet achieved the highest detection accuracy in 11 of the 20 covered categories, with an overall accuracy of 83%.

[0087] Table 6 compares AFENet with 9 other popular methods and benchmark methods on the Microsoft COCO2017 dataset.

[0088]

[0089] Table 6 shows the comparison of the proposed AFENet with nine other popular object detection methods and benchmark methods on the Microsoft COCO2017 dataset. Bold text indicates the maximum value in that column. Although AFENet has a slightly higher number of parameters compared to the benchmark methods, it has a significant advantage in accuracy, especially for small-sized targets. Compared to the DETR method, which is also based on an attention mechanism, AFENet has a substantial advantage in both parameter count and accuracy, with a 4.0% improvement in accuracy (APS) for small targets.

[0090] Table 7 compares AFENet with 9 other popular methods and benchmark methods on the VisDrone2019 dataset.

[0091]

[0092]

[0093] Table 7 shows the comparison between the proposed AFENet and nine other popular object detection methods and benchmark methods on the VisDrone2019 dataset. Bold text indicates the maximum value in that column. As shown in Table 7, compared to the dedicated small object detection model TPH-YOLOv5, AFENet achieves higher detection accuracy with fewer parameters. Compared to the state-of-the-art YOLOv11 and YOLOv12, although AFENet has increased parameters and computational cost, it achieves leading accuracy (AP50 and AP50:95) by 4.8% and 6.3%, respectively.

[0094] Figure 5 Visual ablation experiments were conducted on each module of AFENet. After gradually introducing the MFEM and JRAM modules, AFENet detected small pixel targets such as bicycles, potted plants, and birds that the baseline model could not detect, and the confidence level was further improved.

[0095] exist Figure 6 A comparative visualization experiment was conducted comparing YOLOv12-N, YOLOv8-N, and AFENet. It can be seen that AFENet successfully detected the tricycle on the right side of the first image and the more distant pedestrion category, while the other two models failed to detect them. In the third image, AFENet avoided the false positives of YOLOv8-N regarding pedestrion and truck, and also identified smaller pixel-sized pedestrion and car targets.

[0096] The above experiments demonstrate that AFENet has excellent performance in detecting small and medium-sized objects, especially showing significant performance improvement in small object detection tasks, while maintaining high inference efficiency.

[0097] The present invention also relates to an electronic device, the electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement an attention-driven feature enhancement-based target detection method of the present invention.

[0098] The present invention also relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an attention-driven feature enhancement-based target detection method of the present invention.

[0099] In summary, this invention effectively solves the key problem of feature information loss in small target detection through the synergistic effect of the MFEN and JRAM modules. It can better capture and preserve subtle features in images, significantly improving the recognition ability of small targets and resulting in a marked improvement in the model's detection accuracy in complex scenes. The modular design allows the network to be flexibly adjusted according to actual needs, achieving a balance between detection accuracy and operational efficiency.

[0100] Finally, the improved network structure of this invention has strong generalization ability, and is not only suitable for conventional target detection tasks, but also performs well in small target detection scenarios in professional fields such as UAV aerial photography and medical imaging.

[0101] It should be noted that the above embodiments are only used to illustrate the present invention, but the present invention is not limited to the above embodiments. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A target detection method based on attention-driven feature enhancement, characterized in that, Includes the following steps: S1. Input the preprocessed image into the backbone network built based on YOLOv8; S2. Replace one or more C2f modules in the backbone network with a multi-branch feature preservation and enhancement (MFEM) module to perform feature enhancement and preservation processing. The MFEM module divides the input feature map into two parts: the first part directly retains the original features, and the second part is extracted through multi-branch feature enhancement. Finally, the output features enhanced by multiple branches are concatenated with the original features of the first part to obtain the enhanced feature map. S3. Input the enhanced feature map into the Joint Residual Attention Mechanism (JRAM) module, perform channel and spatial attention weighting in parallel, retain the original features through residual connections, and output the weighted feature map. S4. Input the weighted feature map into the feature pyramid network for multi-scale feature fusion, input the fused features into the detection head, and output the target detection result.

2. The target detection method based on attention-driven feature enhancement according to claim 1, characterized in that, In step S2, the processing procedure of the MFEM module includes: The input feature map is divided into two parts: one part retains the original features, and the other part is enhanced and extracted through multiple branches. Each branch is used to extract and fuse features through the RFACA module and the PConv module in sequence. Finally, the output features of multiple branches are concatenated with the original features, and the number of channels is adjusted through a convolutional layer before output.

3. The target detection method based on attention-driven feature enhancement according to claim 2, characterized in that, The specific processing flow of the MFEM module is as follows: (1) First, adjust the number of channels of the input feature map to fit the dimension of subsequent processing by using 1×1 convolution; (2) Divide the feature map after channel adjustment into branches C1 and C5. C5 directly retains the original features, and C1 is further split into parallel branches C2, C3 and C4. (3) After the attention-weighted features of branches C2, C3 and C4 are extracted by the RFA module, the original features within the branches are retained by the PConv module and the enhanced features output by RFA are fused to obtain the enhanced features of the sub-branches. (4) The output features of C2, C3, and C4 are concatenated with the original features of C1 and C5 through the Concat operation, and then fused by 1×1 convolution and the number of channels is adjusted to obtain the final enhanced features of the MFEM module.

4. The target detection method based on attention-driven feature enhancement according to claim 2, characterized in that, The RFACA module is composed of RFAConv and a coordinate attention mechanism, which is used to provide attention weights to the convolution kernel and capture spatial features.

5. The target detection method based on attention-driven feature enhancement according to claim 1, characterized in that, In step S4, the processing procedure of the JRAM module includes: The input feature maps are fed into the SEA and SA sub-modules for parallel processing. The SEA sub-module is used to calculate the channel dimension weights, and the SA sub-module is used to calculate the spatial dimension weights. Then, the outputs of the two sub-modules are residually concatenated with the original features of the input JRAM module. Finally, the two residuals are concatenated, and the output is adjusted by passing a convolutional layer.

6. The target detection method based on attention-driven feature enhancement according to claim 5, characterized in that, In step S4, the JRAM module is deployed at layers 5, 8, and 11 of the backbone network.

7. The target detection method based on attention-driven feature enhancement according to claim 6, characterized in that, The attention weight parameters of each JRAM module are independently optimized to adapt to the target feature enhancement requirements at the corresponding resolution.

8. An electronic device, the electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the target detection method based on attention-driven feature enhancement as described in any one of claims 1-8.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements an attention-driven feature enhancement-based target detection method as described in any one of claims 1-8.

Images