An epidemic prevention scene target detection method

Abstract

The application discloses an epidemic prevention scene target detection method, which first uses a deep convolutional neural network in a YOLO series basic network and FPN technology for feature extraction and cross-scale processing on an input image to generate three layers of feature maps; then uses a feature fusion module to obtain three layers of fusion enhanced feature maps by using the features of each layer and combining the cross-scale feature mapping of other layers; subsequently uses a feature denoising module to generate three layers of denoised feature maps by using a channel attention mechanism, convolution operation and Softmax processing, and uses a salient feature map and a binary map to calculate attention loss; finally uses a YOLO series prediction module to perform classification and coordinate regression, and displays output after non-maximum suppression processing. The application fully utilizes the multi-layer feature information of the target in the image, effectively suppresses the noise information existing in the background information, has obvious effect on improving detection precision, and has good application prospect and economic benefits.

Description

Technical Field

[0001] This invention belongs to the field of natural image target detection, and relates to a target detection method in an epidemic prevention scenario, particularly a target detection method based on feature denoising in an epidemic prevention scenario. Background Technology

[0002] Early object detection primarily relied on traditional algorithms, with feature extraction largely done manually. This approach had significant limitations and poor robustness, and the network performance could not meet practical needs. Since the introduction of AlexNet, the first convolutional neural network, in 2012, the integration of object detection with convolutional neural networks ushered in an unprecedented new era. Subsequently, the R-CNN algorithm was proposed in 2014, and object detection has developed rapidly in this direction ever since. Object detection algorithms in deep learning can be broadly categorized into two types based on their detection principles: two-stage detection and one-stage detection. Two-stage algorithms use generated bounding boxes as a foundation, employing extracted features to produce refined bounding boxes and predict the object's category. They are characterized by high accuracy but slow speed. One-stage algorithms combine bounding box generation, classification, and bounding box refinement into a single process. They are characterized by high speed but slightly lower accuracy. The YOLO series of algorithms is a classic example of a single-stage algorithm.

[0003] Because multi-target detection in epidemic prevention scenarios requires near real-time speeds, target detection networks based on single-stage algorithms are more suitable for this application. However, the targets in these scenarios vary in size, encompassing large, medium, and small targets, and the complex background information makes it easy for the edges of targets such as crowds to be confused with the background. This means that existing detection algorithms cannot meet the actual detection requirements. Therefore, it is necessary to improve the target detection network to enhance detection accuracy. The core of this improvement lies in removing background noise and enhancing the feature information contained in the target to overcome problems such as complex backgrounds and insufficient object feature extraction. The key technology lies in effectively combining target feature information and using it to generate accurate prediction results. Summary of the Invention

[0004] In view of the above-mentioned prior art, the technical problem to be solved by the present invention is to provide a target detection method based on feature denoising in epidemic prevention scenarios, which can effectively improve the detection accuracy of targets in natural images.

[0005] To address the aforementioned technical problems, the present invention provides a target detection method for epidemic prevention scenarios, comprising the following steps:

[0006] Step 1: Use the deep convolutional neural network in the YOLO series of basic networks as the backbone network to extract features from the input image. Then use the feature pyramid technique to horizontally concatenate the three feature maps from top to bottom to output three feature maps {F2, F3, F4}.

[0007] Step 2: Using the feature fusion module, the three-layer feature maps {F2, F3, F4} obtained in Step 1 are used to apply the features of each layer and combine them with the cross-scale feature mappings of other layers to obtain the three-layer fused enhanced feature maps {M2, M3, M4}.

[0008] Step 3: The three-layer fused enhanced feature maps {M2, M3, M4} obtained in Step 2 are processed by the feature denoising module. The channel attention mechanism and convolution operation are used sequentially to generate salient feature maps. After Softmax processing, the maps are multiplied by the input to obtain the three-layer denoised feature maps {D2, D3, D4}. The attention loss is also calculated.

[0009] Step 4: Using the three-layer denoised feature maps {D2, D3, D4} and attention loss obtained in Step 3, perform classification and coordinate regression using the YOLO series prediction modules, and display the prediction results after non-maximum suppression processing.

[0010] Furthermore, in step 2, the three-layer feature maps {F2, F3, F4} obtained in step 1 are processed using the feature fusion module. The features of each layer are used and combined with cross-scale feature mappings from other layers to obtain the three-layer fused enhanced feature maps {M2, M3, M4}, including:

[0011] Step 2.1: Use a 1×1 convolution operation to change the channel dimension of the three-layer feature map {F2, F3, F4} output from Step 1, and output the compressed three-layer feature map {P2, P3, P4}.

[0012] Step 2.2: For the feature maps P2 and P3 generated in Step 2.1, both are downsampled to make them the same size as the P4 layer, and then stitched together in the channel dimension. They are then added to the feature map F4 obtained in Step 1 to generate the fused and enhanced feature map M4.

[0013] Step 2.3: Apply downsampling operation to feature map P2 generated in step 2.1 and upsampling operation to P4 to make them the same size as layer P3, and then stitch them together in the channel dimension. Then add them to feature map F3 obtained in step 1 to generate fused enhanced feature map M3.

[0014] Step 2.4: Upsample layers P3 and P4 in the three-layer feature map generated in Step 2.1 to make them the same size as layer P2, and then stitch them together in the channel dimension. Then add them to the feature map F2 obtained in Step 1 to generate the fused enhanced feature map M2.

[0015] Furthermore, in step 3, the three-layer fused enhanced feature maps {M2, M3, M4} obtained in step 2 are processed using the feature denoising module. Channel attention and convolution operations are used sequentially to generate salient feature maps, which are then processed by Softmax and multiplied by the input to obtain the three-layer denoised feature maps {D2, D3, D4}. The attention loss is calculated as follows:

[0016] Step 3.1: Input the fused enhanced feature map M2 obtained in Step 2 into the channel attention module and use global average pooling, one-dimensional convolution operation and activation function to calculate local cross-channel interaction in sequence, and multiply the result with the input feature map M2 to obtain the channel interaction feature map;

[0017] Step 3.2: Perform a convolution operation on the output of Step 3.1 to generate a salient feature map with foreground scores. Then, process it using the Softmax operation and multiply it with the fused and enhanced feature map M2 to obtain the denoised feature map D2.

[0018] Step 3.3: Perform pixel-level cross-entropy loss calculation on the salient feature map obtained in Step 3.2 and the binary map obtained from the ground truth bounding box to obtain the loss Loss2;

[0019] Step 3.4: Repeat steps 3.1, 3.2, and 3.3 for the fused and enhanced feature maps M3 and M4 obtained in step 2, respectively, to obtain the denoised feature maps D3 and D4 and their corresponding losses Loss3 and Loss4;

[0020] Step 3.5: Calculate the sum of Loss2, Loss3, and Loss4 obtained in Steps 3.3 and 3.4, and use it as the attention loss of the feature denoising module.

[0021] Furthermore, in step 4, the three-layer denoised feature maps {D2, D3, D4} and attention loss obtained in step 3 are used for classification and coordinate regression using YOLO prediction modules. After non-maximum suppression processing, the prediction results are displayed as follows:

[0022] S4.1: For each layer of denoised feature map {D2, D3, D4}, divide it into fixed regions of the same size. For the target appearing in each fixed region, predict the position of multiple prediction boxes and calculate the corresponding confidence information.

[0023] S4.2: Set a threshold for multiple predicted boxes for each target, perform non-maximum suppression to remove redundant predicted boxes with poor scores, and output the final detection result.

[0024] The beneficial effects of this invention are:

[0025] This invention proposes a feature-based denoising target detection method for epidemic prevention scenarios. First, the input image is processed using a deep convolutional neural network (DCNN) and FPN technology from the YOLO series of basic networks for feature extraction and cross-scale processing, generating a three-layer feature map. Then, a feature fusion module is used to apply the features of each layer of the feature map and combine them with cross-scale feature mappings from other layers to obtain a three-layer fused enhanced feature map. Subsequently, a feature denoising module is used to generate a three-layer denoised feature map using channel attention, convolution operations, and softmax processing, and attention loss is calculated using salient feature maps and binary maps. Finally, the YOLO series of prediction modules are used for classification and coordinate regression, and the output is displayed after non-maximum suppression. The proposed algorithm fully utilizes the multi-layer feature information of the target in the image and effectively suppresses noise information present in the background, significantly improving detection accuracy. This method has good application prospects and economic benefits. Attached Figure Description

[0026] Figure 1 This is an overall schematic diagram of the target detection method based on feature denoising in the epidemic prevention scenario provided by the embodiments of the present invention;

[0027] Figure 2 This is a schematic diagram of the feature fusion module network structure provided in an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of the feature denoising module network structure provided in this embodiment of the invention, taking the fused enhanced feature map M2 as an example as the input.

[0029] Figure 4 This is a schematic diagram of the five target detection categories provided in the embodiments of the present invention;

[0030] Figure 5 This is a schematic diagram of the detection results of the feature-based noise reduction target detection method in the epidemic prevention scenario provided by the embodiments of the present invention. Figure 1 .

[0031] Figure 6 This is a schematic diagram of the detection results of the feature-based noise reduction target detection method in the epidemic prevention scenario provided by the embodiments of the present invention. Figure 2 . Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] This invention first uses the backbone network of the YOLO series of basic networks and FPN technology for feature extraction and cross-scale processing. Then, it uses feature fusion and feature denoising modules to fuse the semantic and localization information of the target and remove background noise, thereby improving the low accuracy of single-stage algorithms. To achieve the above objectives, the method of this invention is as follows:

[0034] (1) Use the deep convolutional neural network in the YOLO series of basic networks as the backbone network to extract features from the input image. Then use the feature pyramid (FPN) technique to horizontally stitch the three feature maps from top to bottom to output three feature maps {F2, F3, F4}.

[0035] (2) Using the feature fusion module, the three-layer feature maps {F2, F3, F4} obtained in (1) are used to apply the features of each layer and combine them with the cross-scale feature mapping of other layers to obtain the three-layer fused enhanced feature maps {M2, M3, M4}.

[0036] (3) The three-layer fused enhanced feature map {M2, M3, M4} obtained in (2) is used with the feature denoising module to generate a significant feature map by sequentially using the channel attention mechanism and convolution operation. Then, after Softmax processing, it is multiplied with the input to obtain the three-layer denoised feature map {D2, D3, D4}, and the attention loss is calculated.

[0037] (4) The three-layer noise reduction feature map {D2, D3, D4} and attention loss obtained in (3) are classified and regressed using the YOLO series prediction module. The prediction results are displayed after non-maximum suppression is applied.

[0038] The feature fusion module uses the features of each layer's feature map and combines them with cross-scale feature maps from other layers to obtain a three-layer fused enhanced feature map. The specific training and detection process includes:

[0039] (2-a): Use a 1×1 convolution operation to change the channel dimension of the three-layer feature map {F2, F3, F4} output by (1) and output the compressed three-layer feature map {P2, P3, P4}.

[0040] (2-b): For the feature maps P2 and P3 generated in (2-a), both are downsampled to make them the same size as the P4 layer, and then stitched together in the channel dimension. They are then added to the feature map F4 obtained in step 1 to generate the fused and enhanced feature map M4.

[0041] (2-c): Apply downsampling operation to feature map P2 generated in (2-a) and upsampling operation to P4 to make it consistent with the size of layer P3, and then stitch them together in the channel dimension. Then add them to feature map F3 obtained in step 1 to generate fused enhanced feature map M3.

[0042] (2-d): For (2-a), the P3 and P4 layers in the three-layer feature map generated in step 2.1 are upsampled to make them the same size as the P2 layer, and then spliced ​​in the channel dimension. Then, they are added to the feature map F2 obtained in step 1 to generate the fused enhanced feature map M2.

[0043] The feature denoising module processes the three-layer fusion enhancement feature map of the feature fusion module, thereby suppressing unwanted noise information mixed in with the feature map. The specific training and detection process includes:

[0044] (3-a): The fusion enhancement feature map M2 obtained in (2) is input into the channel attention module and global average pooling, one-dimensional convolution operation and activation function are used in sequence to calculate local cross-channel interaction, and the result is multiplied with the input feature map M2 to obtain the channel interaction feature map;

[0045] (3-b): Perform a convolution operation on the output of (3-a) to generate a salient feature map with foreground score. Then, process it using the Softmax operation and multiply it with the fused enhanced feature map M2 to obtain the denoised feature map D2.

[0046] (3-c): Calculate the pixel-level cross-entropy loss by comparing the salient feature map obtained in (3-b) with the binary map derived from the ground truth bounding box, and obtain the loss Loss2.

[0047] (3-d): Repeat steps (3-a), (3-b), and (3-c) for the fusion enhancement feature maps M3 and M4 obtained in (2) to obtain the denoising feature maps D3 and D4 and their corresponding losses Loss3 and Loss4.

[0048] (3-e): The sum of Loss2, Loss3, and Loss4 obtained from (3-a), (3-b), (3-c), and (3-d) is used as the attention loss of the feature denoising module.

[0049] Combination Figure 1 The specific implementation steps of the present invention are as follows, with specific parameters provided in the examples:

[0050] S1. The dataset, after being filtered and supplemented with categories, is randomly divided into training, validation, and test sets in a ratio of 5:3:2. Then, the images in the training set are input into the backbone network of the YOLO series basic network and FPN technology for feature extraction and cross-scale processing to generate three-layer feature maps {F2, F3, F4}.

[0051] The deep convolutional neural network in the YOLO series serves as the backbone network. Its function is to extract features from images that are conducive to target classification and localization through many convolutional layers of different sizes. As the network learns autonomously based on a large amount of data containing target information, it changes the network parameters through forward and backward propagation, thereby gradually improving the network's detection performance.

[0052] FPN technology primarily addresses the multi-scale problem in object detection. By altering the connections between feature layers through a simple top-down, horizontally connected network structure, it obtains more contextual information without significantly increasing the computational load of the original model.

[0053] S2. Using the feature fusion module, apply the features of each layer in the feature maps of {F2, F3, F4}, and combine them with the cross-scale feature maps of other layers to obtain the three-layer fused enhanced feature maps {M2, M3, M4}, as shown below. Figure 2 As shown, the specific implementation steps are as follows:

[0054] S21: Use a 1×1 convolution operation to change the channel dimension of the three-layer feature map {F2, F3, F4} output by S1, and output the compressed three-layer feature map {P2, P3, P4}.

[0055] S22: For feature maps P2 and P3 generated in S21, downsampling is used to make them consistent with the size of layer P4, and they are spliced ​​in the channel dimension. Then they are added to feature map F4 obtained in S1 to generate fused and enhanced feature map M4.

[0056] S23: Apply downsampling operation to feature map P2 generated by S21 and upsampling operation to P4 to make it consistent with the size of layer P3, and then stitch them together in the channel dimension. Then add them to feature map F3 obtained by S1 to generate fused enhanced feature map M3.

[0057] S24: Upsample layers P3 and P4 in the three-layer feature map generated in S21 to make them the same size as layer P2, and then stitch them together in the channel dimension. Then add them to the feature map F2 obtained in S1 to generate the fused enhanced feature map M2.

[0058] In convolutional neural networks, as the number of layers increases, the semantic information represented by the acquired feature maps also changes. Deeper feature maps often contain more semantic features, which is beneficial for classification tasks, while shallower feature maps contain more image features, which is beneficial for localization. However, although deep feature maps can respond to semantic features, their small size means they contain limited geometric information, which is detrimental to object detection. Shallow feature maps, while containing more geometric information, lack semantic features, making them less suitable for image classification. This problem is even more pronounced in small object detection. Therefore, merging deep and shallow features can simultaneously meet the needs of object detection and image classification.

[0059] In subsequent prediction processes, this operation not only utilizes the feature information of the current layer, but also takes into account the context information of the upper and lower layers. The resulting new feature map has a richer amount of information. Therefore, by utilizing features from both high and low layers, this approach can address the challenges of large variations in target size and inter-class similarity, thereby achieving high-quality target detection.

[0060] S3. The feature denoising module is used to process the three-layer fusion enhancement feature map {M2, M3, M4} of the feature fusion module, so that the useless noise information mixed in the feature map is suppressed. Figure 3 The diagram shows the network structure of the feature denoising module, with the input being the fused enhanced feature map M2. The specific training and detection process includes:

[0061] S31: The M2 layer of the three-layer fused enhanced feature map {M2, M3, M4} obtained in S2 is input into the channel attention module. First, global average pooling is used to compress the dimension of the feature map, generating a vector of size 1×1×C containing all feature information. Then, a one-dimensional convolution operation of size k=5 is used to calculate the output channel attention weights by utilizing the channel interaction between a certain channel and its k neighboring channels. After processing by the activation function, a channel attention weight vector is generated, which is then multiplied with the input feature map M2 to obtain the channel interaction feature map.

[0062] When calculating channel attention weights, using all channel information for every position reduces computational efficiency because channel interactions in some channels are not related to all channels. Furthermore, using one-dimensional convolution avoids the negative impact of dimensionality reduction on weight calculation and provides good cross-channel information acquisition capabilities. Appropriate cross-channel interactions can significantly reduce model complexity while maintaining performance, achieving noticeable performance gains through adjustments to a few parameters. The formula for calculating channel attention weights is shown below:

[0063]

[0064] in Input feature map; This is a global average pooling operation; This is a one-dimensional convolution operation of size k×k; For activation functions; This represents the channel attention weight.

[0065] S32: Perform a convolution operation on the channel interaction feature map output by S31 to generate a salient feature map with foreground score. Then, use the Softmax operation to process it to reduce the noise information contained in the feature map. Then multiply it with the fused enhanced feature map M2 to obtain the denoised feature map D2.

[0066] After passing through convolutional layers, each layer of fused and enhanced feature maps generates a salient feature map with a foreground score. The subsequent Softmax processing, due to the function mapping, ensures that the output is between 0 and 1, thus enhancing target features and suppressing background noise. Furthermore, the subsequent multiplication with the fused and enhanced feature map output from S2 prevents non-target information from being completely eliminated but rather partially suppressed, thereby preserving contextual information and improving network robustness.

[0067] S33: Perform pixel-level cross-entropy loss calculation on the salient feature map obtained in S32 and the binary map obtained from the ground truth bounding box to obtain the loss Loss2.

[0068] This attention loss function can be used in the forward and backward propagation processes of the prediction module in the YOLO series basic network for training. The loss function is used to measure the difference between the actual variable value and the predicted value. The smaller the loss value, the more accurate the prediction. Therefore, the loss function value can be minimized by training data, thereby training a network structure with good detection performance.

[0069] S34: Repeat S31, S32, and S33 for the fused and enhanced feature maps M3 and M4 obtained in S2 to obtain denoised feature maps D3 and D4 and their corresponding losses Loss3 and Loss4.

[0070] S35: The sum of Loss2, Loss3, and Loss4 obtained from S31, S32, S33, and S34 is used as the attention loss of the feature denoising module.

[0071] S4. The three-layer denoised feature maps {D2, D3, D4} after feature denoising are combined with the attention loss using YOLO series prediction modules for classification and coordinate regression. After non-maximum suppression processing, the detected mask-wearing method is output. The specific training and detection process includes:

[0072] S41: For each layer of denoised feature map {D2, D3, D4}, divide it into fixed regions of the same size. For targets appearing in each fixed region, predict the positions of multiple prediction boxes and calculate the corresponding confidence information.

[0073] S42: Set a threshold for multiple predicted boxes for each target, perform non-maximum suppression to remove redundant predicted boxes with poor scores, and output the final detection result.

[0074] Figure 4 The document provides diagrams for five categories: non-motorized vehicles (riders), pedestrians (pedestrians), epidemic prevention workers (EP_workers), microbuses, and cars (Cars). Figure 5 , 6 The results are shown for five target categories. It can be observed from the figure that, apart from the very small targets in the image, most targets are accurately identified. Even if the background where the pedestrian appears is complex, it can still be accurately identified.

[0075] Table 1 lists the detection results for four mask-wearing methods in this embodiment. The image dataset was selected based on the publicly available datasets BITVehicle_Dataset and WiderPerson, with data suitable for the scenario added and the category of epidemic prevention personnel added. Then, the LabelImg tool was used to create XML files storing the target location and category, and all data types were converted to the PASCAL VOC dataset format, and then to the format required by the YOLO algorithm. The dataset was divided into training, validation, and test sets in a 5:3:2 ratio for model training, validation, and testing. The results in the table show that the feature-based denoising target detection method used in this example for epidemic prevention scenarios has high detection accuracy for the five categories.

[0076] Table 1. Comparison of Detection Results for Four Types of Mask Wearing in This Example

[0077] Category Pedestrian Epidemic prevention worker Car Small passenger car Non-motor vehicle mAP AP 0.758 0.898 0.947 0.920 0.628 0.830

[0078] In summary, the feature-based denoising object detection method in epidemic prevention scenarios effectively improves the accuracy and detection performance of natural images, demonstrating strong practical applicability. The proposed feature-based denoising object detection method for epidemic prevention scenarios can effectively fuse semantic and positional information of targets at different scales and suppress noise mixed in the feature map, thereby improving the detection performance of targets that are difficult to identify in complex backgrounds. Furthermore, this method is not limited to the five categories mentioned above; after modifying the dataset and adjusting the parameters, it can also be applied to other categories in natural images.

[0079] The above describes the basic principles, main features, and practical advantages of a feature-based noise reduction target detection method in an epidemic prevention scenario. Those skilled in the art should understand that the above embodiments are only for helping to understand the method and core ideas of the present invention and do not limit the present invention. At the same time, based on the ideas of this application, there will be changes in specific implementation methods and application scope, and these changes all fall within the protection scope of the present invention.

Claims

1. A target detection method in an epidemic prevention scenario, characterized in that, Includes the following steps: Step 1: Use the deep convolutional neural network in the YOLO series of basic networks as the backbone network to extract features from the input image. Then use the feature pyramid technique to horizontally concatenate the three feature maps from top to bottom to output three feature maps {F2, F3, F4}. Step 2: Using the feature fusion module, the three-layer feature maps {F2, F3, F4} obtained in Step 1 are used to apply the features of each layer and combine them with the cross-scale feature mappings of other layers to obtain the three-layer fused enhanced feature maps {M2, M3, M4}. Step 3: The three-layer fused enhanced feature map {M2,M3,M4} obtained in Step 2 is processed by the feature denoising module. The channel attention mechanism and convolution operation are used to generate salient feature maps in sequence. Then, after Softmax processing, it is multiplied with the input to obtain the three-layer denoised feature map {D2,D3,D4}, and the attention loss is calculated. Step 3.1: Input the fused enhanced feature map M2 obtained in Step 2 into the channel attention module and use global average pooling, one-dimensional convolution operation and activation function to calculate local cross-channel interaction in sequence, and multiply the result with the input feature map M2 to obtain the channel interaction feature map; Step 3.2: Perform a convolution operation on the output of Step 3.1 to generate a salient feature map with foreground scores. Then, process it using the Softmax operation and multiply it with the fused and enhanced feature map M2 to obtain the denoised feature map D2. Step 3.3: Perform pixel-level cross-entropy loss calculation on the salient feature map obtained in Step 3.2 and the binary map obtained from the ground truth bounding box to obtain the loss Loss2; Step 3.4: Repeat steps 3.1, 3.2, and 3.3 for the fused and enhanced feature maps M3 and M4 obtained in step 2, respectively, to obtain the denoised feature maps D3 and D4 and their corresponding losses Loss3 and Loss4; Step 3.5: Calculate the sum of Loss2, Loss3, and Loss4 obtained in Steps 3.3 and 3.4, and use it as the attention loss of the feature denoising module; Step 4: Using the three-layer denoised feature maps {D2,D3,D4} and attention loss obtained in Step 3, perform classification and coordinate regression using the YOLO series prediction modules, and display the prediction results after non-maximum suppression processing.

2. The target detection method in an epidemic prevention scenario according to claim 1, characterized in that: Step 2, which involves using the feature fusion module to process the three-layer feature maps {F2, F3, F4} obtained in Step 1, uses the features of each layer and combines them with the cross-scale feature mappings of other layers to obtain the three-layer fused enhanced feature maps {M2, M3, M4}, includes: Step 2.1: Use a 1×1 convolution operation to change the channel dimension of the three-layer feature map {F2,F3,F4} output from Step 1, and output the compressed three-layer feature map {P2,P3,P4}. Step 2.2: For the feature maps P2 and P3 generated in Step 2.1, both are downsampled to make them the same size as the P4 layer, and then stitched together in the channel dimension. They are then added to the feature map F4 obtained in Step 1 to generate the fused and enhanced feature map M4. Step 2.3: Apply downsampling operation to feature map P2 generated in step 2.1 and upsampling operation to P4 to make them the same size as layer P3, and then stitch them together in the channel dimension. Then add them to feature map F3 obtained in step 1 to generate fused enhanced feature map M3. Step 2.4: Upsample layers P3 and P4 in the three-layer feature map generated in Step 2.1 to make them the same size as layer P2, and then stitch them together in the channel dimension. Then add them to the feature map F2 obtained in Step 1 to generate the fused enhanced feature map M2.

3. The target detection method in an epidemic prevention scenario according to claim 1, characterized in that: Step 4 describes using the three-layer denoised feature maps {D2, D3, D4} and attention loss obtained in Step 3 for classification and coordinate regression using YOLO prediction modules. After non-maximum suppression processing, the prediction results are displayed as follows: S4.1: For each layer of denoised feature map {D2,D3,D4}, divide it into fixed regions of the same size. For targets appearing in each fixed region, predict the positions of multiple prediction boxes and calculate the corresponding confidence information. S4.2: Set a threshold for multiple predicted boxes for each target, perform non-maximum suppression to remove redundant predicted boxes with poor scores, and output the final detection result.

Images