Multi-scale target detection method based on YOLOv5s

Abstract

The invention provides a multi-scale target detection method based on YOLOv5s, and the method is characterized in that the method comprises the following steps: S1, obtaining a data set; s2, preprocessing the data set; s3, constructing an MSFS module with cross-scale perception; s4, constructing a TFL module used for position positioning; s5, the MSFS module and the TFL module are embedded into Neck, and an optimized multi-scale target detection model based on YOLOv5s is obtained; and S6, inputting the preprocessed data set into the optimized YOLOv5s-based multi-scale target detection model to obtain a final data set. According to the comprehensive design, the module can comprehensively consider the channel and spatial relationship, so that the CA mixed domain attention mechanism can effectively solve the problem that the position and spatial characteristics of a detected target and the position characteristics of an optimization model detected target are not obvious.

Description

Technical Field

[0001] The present application relates to the technical field of multi-scale target detection in autonomous driving, and specifically to a multi-scale target detection method based on YOLOv5s. Background Art

[0002] In the autonomous driving environment, there are many types of targets, including cars, pedestrians, motorcycles, etc., and their scales and appearances vary greatly. At the same time, the background environment of the target to be detected is complex and changeable, requiring the target detection model to have a high degree of adaptability. In addition, since distant targets are relatively small and nearby targets are relatively large, and there are a large number of targets in complex traffic road scenes, occlusion may occur between targets, resulting in a surge in the number of small targets and occluded targets, which poses a challenge to the accuracy of target detection.

[0003] In the task of object detection, the multi-scale problem has always been the key of research. The multi-scale problem refers to the situation where the objects to be detected have significant differences in size. Distant objects are small, while nearby objects are large, or pedestrians and motorcycles are small in size relative to vehicles. In the task of object detection in complex scenes, due to the large number of objects and the different sizes of distant and near objects, the multi-scale problem is significant, which affects the accuracy of object detection.

[0004] Mao Zhen et al. proposed an improved multi-scale target detection algorithm based on YOLOv5s. The CBAM module is integrated between the YOLOv5s backbone network and the Neck network to enhance the model's ability to extract important feature information of multi-scale targets. Qian Wu, Wang Guozhong et al. proposed a multi-scale YOLOv5 traffic light detection method. First, the composite data enhancement method is used to enhance the model input to increase the complexity of the input; then, multi-scale training is designed to replace fixed-scale training to balance the learning ability of the model; finally, a multi-scale feature fusion network is constructed to fuse 4x, 8x, 16x and 32x downsampling information to establish a multi-scale detection layer. In order to enhance the feature fusion capability, far-jump links are introduced to transmit different levels of information, which directly improves the model's detection capability for small targets. Wang Shihang uses YOLOv5s as the baseline model and proposes a lightweight multi-scale target detection model YOLOs-SK (YOLOsmall-), which improves the model's detection capability for multi-scale targets while ensuring that the model complexity remains basically unchanged. Duan Shengtian proposed a multi-scale traffic sign recognition algorithm model CBAM-BiYOLOv5 to reduce the missed detection rate of small targets. Zhu Ningke, Zhang Shudi, Wang Hanwen, etc. proposed an improved YOLOv5 algorithm based on multi-scale fusion to enhance the detection ability of the network. Hui Kanghua, Yang Wei, Liu Haohan, etc. proposed an enhanced multi-scale target detection (EM-YOLOv5) method based on YOLOv5. Through the Kmeans++ clustering algorithm, a multi-scale initialization anchor frame adapted to the current detection scenario is obtained, making it easier for the network to capture targets of different scales; multiple parallel convolution branches of different scales are added to the Bottleneck structure to fuse multi-scale feature information while retaining the original feature information, thereby enhancing the global perception ability of the model. Summary of the invention

[0005] In order to solve the above technical problems, a multi-scale target detection method based on YOLOv5s is proposed, which includes the following steps:

[0006] S1, dataset acquisition;

[0007] S2, preprocessing the data set;

[0008] S3, build an MSFS module with cross-scale perception;

[0009] S4, constructing a TFL module for position positioning;

[0010] S5, embed the MSFS module and the TFL module into Neck to obtain the optimized multi-scale target detection model based on YOLOv5s;

[0011] S6. Input the preprocessed data set into the optimized multi-scale object detection model based on YOLOv5s to obtain the final data set.

[0012] The above technical solution proposes a multi-scale target detection method based on YOLOv5s. By constructing the MSFS module and the TFL module and embedding Neck, the target detection model is optimized, which can achieve effective detection of multi-scale targets and obtain the final data set. It provides a systematic multi-scale target detection method and solves the problem of how to effectively detect multi-scale targets.

[0013] Specifically, the MSFS module uses dilated convolutions with different dilation rates to extract features of targets of different scales, and combines the spatial pyramid structure to construct the MSFS module.

[0014] The above technical solution clarifies that the MSFS module uses dilated convolutions with different dilation rates and spatial pyramid structures to extract features of targets of different scales, improves the accuracy and comprehensiveness of feature extraction, and solves the problem of how to accurately extract features of targets of different scales to meet the needs of multi-scale target detection.

[0015] Specifically, the MSFS module consists of five parallel branches and dilated convolutions with different dilation rates.

[0016] The above technical solution specifically points out that the MSFS module includes five parallel branches and dilated convolutions with different dilation rates, which provides a specific structural basis for realizing multi-scale feature extraction, further clarifies the structural composition of the MSFS module, and solves the specific implementation problems of module construction.

[0017] Specifically, the five parallel branches in the MSFS module are used to input the received preprocessed data sets into one-dimensional convolution, dilated convolution with a dilation rate of 6, dilated convolution with a dilation rate of 8, dilated convolution with a dilation rate of 12, and maximum pooling, respectively.

[0018] The above technical solution details how the five parallel branches process the preprocessed dataset, which helps to extract multi-scale features more accurately and solves the problem of how to input the dataset into the MSFS module for effective processing.

[0019] Specifically, the S3 specific method is as follows:

[0020] The preprocessed data sets are input into five parallel branches respectively; the preprocessed data sets input into the first branch enter the one-dimensional convolution to save the key features in the preprocessed data sets; the preprocessed data sets input into the second to fifth branches enter the first dilated convolution, the second dilated convolution, the third dilated convolution with dilation rate and the maximum pooling respectively.

[0021] The above technical solution describes the specific processing flow of the preprocessed data set in S3 in five parallel branches in the MSFS module, ensures the preservation of key features and the extraction of multi-scale features, and solves the specific steps and process problems of data processing in the MSFS module.

[0022] Specifically, the S3 method also includes: the preprocessed data set is subjected to a first dilated convolution to output a first feature map, the first feature map is input into a second dilated convolution to output a second feature map, and the second feature map is input into a third dilated convolution to output a third feature map.

[0023] The above technical solution illustrates the feature map output process of the preprocessed dataset after dilated convolution, which helps to gradually extract and fuse multi-scale features and solves the problem of gradual extraction and fusion of multi-scale features in the MSFS module.

[0024] Specifically, the TFL module consists of two ConvBnSiLU modules, a CA mixed-domain attention mechanism, and a residual structure. Specifically, the ConvBnSiLU module consists of a 2D convolutional layer, a BN layer, and a SiLU activation function.

[0025] Specifically, the S4 specific method is as follows:

[0026] The feature map in the input TFL module is mapped to a more semantic representation space through the first ConvBnSiLU module. The mapped feature map is output to the second ConvBnSiLU module through the CA mixed domain attention mechanism. The second ConvBnSiLU module performs a convolution operation on the feature map fused with the feature map output by the CA mixed domain attention mechanism. Specifically, the CA mixed domain attention mechanism is used to assign weights of different spatial positions to the input feature map.

[0027] The above technical solution points out that the role of the CA mixed domain attention mechanism is to give weights to different spatial positions of the feature map, which helps to locate the position more accurately and solves the problem of how to improve the accuracy of position positioning through the attention mechanism.

[0028] Compared with the prior art, the beneficial results of this application are:

[0029] By cleverly integrating the output feature maps of these five branches, the MSFS module achieves comprehensive utilization of features at different scales. This multi-scale feature extraction mechanism provides more comprehensive and accurate information for the detection model, significantly improving the performance of the model in multi-scale object detection tasks;

[0030] This enables the module to comprehensively consider the channel and spatial relationship, so that the CA mixed domain attention mechanism can effectively solve the position and spatial characteristics of the detection target, and optimize the model to detect the problem of unclear target position characteristics. It cooperates with the MSFS cross-scale perception module to improve the robustness and accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated into and constitute a part of this specification. The accompanying drawings illustrate the embodiments and are used together with the description to explain the principles of the present application. It will be easy to recognize other embodiments and many expected advantages of the embodiments because they become better understood by reference to the following detailed description. The elements of the drawings are not necessarily to scale with each other. The same reference numerals refer to corresponding similar parts.

[0032] Figure 1 It is a flowchart of a multi-scale target detection method based on YOLOv5s according to an embodiment of the present application;

[0033] Figure 2 is a schematic diagram of the structure of an MSFS module of a multi-scale target detection method based on YOLOv5s according to an embodiment of the present application;

[0034] Figure 3 is a schematic diagram of the TFL module structure of a multi-scale target detection method based on YOLOv5s according to an embodiment of the present application;

[0035] Figure 4 It is a comparison diagram between a multi-scale target detection method based on YOLOv5s according to an embodiment of the present application and YOLOv5s, YOLOv5s+SPP and YOLOv5s+SPPF;

[0036] Figure 5 It is a performance comparison diagram between a multi-scale target detection method based on YOLOv5s according to an embodiment of the present application and YOLOv5s, YOLOv5s+SPP and YOLOv5s+SPPF;

[0037] Figure 6 This is a comparison chart of the detection performance differences between a multi-scale target detection method based on YOLOv5s according to an embodiment of the present application and the original model YOLOv5s. DETAILED DESCRIPTION

[0038] In the following detailed description, reference is made to the accompanying drawings, which form a part of the detailed description and are shown by illustrative specific embodiments in which the present application can be practiced. To this end, directional terms, such as "top", "bottom", "left", "right", "up", "down", etc., are used with reference to the orientation of the figures described. Because the components of the embodiments can be positioned in several different orientations, directional terms are used for the purpose of illustration and are by no means limiting. It should be understood that other embodiments may be utilized or logical changes may be made without departing from the scope of the present application. Therefore, the following detailed description should not be adopted in a limiting sense, and the scope of the present application is defined by the appended claims.

[0039] like Figures 1 to 3 As shown, a multi-scale target detection method based on YOLOv5s includes the following steps:

[0040] S1, dataset acquisition;

[0041] S2, preprocessing the data set;

[0042] S3, build an MSFS module with cross-scale perception;

[0043] S4, constructing a TFL module for position positioning;

[0044] S5, embed the MSFS module and the TFL module into Neck to obtain the optimized multi-scale target detection model based on YOLOv5s;

[0045] S6. Input the preprocessed data set into the optimized multi-scale object detection model based on YOLOv5s to obtain the final data set.

[0046] The MSFS module uses dilated convolutions with different dilation rates to extract features of targets of different scales, and combines the spatial pyramid structure to build the MSFS module.

[0047] The MSFS module consists of five parallel branches and dilated convolutions with different dilation rates.

[0048] The five parallel branches in the MSFS module are used to input the received preprocessed data sets into one-dimensional convolution, dilated convolution with a dilation rate of 6, dilated convolution with a dilation rate of 8, dilated convolution with a dilation rate of 12, and maximum pooling, respectively.

[0049] The specific method of S3 is as follows:

[0050] The preprocessed data sets are input into five parallel branches respectively; the preprocessed data sets input into the first branch enter the one-dimensional convolution to save the key features in the preprocessed data sets; the preprocessed data sets input into the second to fifth branches enter the first dilated convolution, the second dilated convolution, the third dilated convolution with dilation rate and the maximum pooling respectively.

[0051] The specific method of S3 also includes: the preprocessed data set is subjected to the first dilated convolution to output a first feature map, the first feature map is input into the second dilated convolution to output a second feature map, and the second feature map is input into the third dilated convolution to output a third feature map.

[0052] The TFL module consists of two ConvBnSiLU modules, a CA mixed-domain attention mechanism, and a residual structure. The ConvBnSiLU module consists of a 2D convolutional layer, a BN layer, and a SiLU activation function.

[0053] The specific method of S4 is as follows:

[0054] The feature map in the input TFL module is mapped to a more semantic representation space through the first ConvBnSiLU module. The mapped feature map is output to the second ConvBnSiLU module through the CA mixed domain attention mechanism. The second ConvBnSiLU module performs a convolution operation on the feature map fused with the feature map output by the CA mixed domain attention mechanism. The CA mixed domain attention mechanism is used to assign weights of different spatial positions to the input feature map.

[0055] More specifically, Figure 2 As shown in the figure, when the convolution kernel of the ordinary convolution slides on the feature map, the input value of each position will be calculated with the corresponding weight in the convolution kernel. The dilated convolution introduces evenly distributed holes in the convolution kernel, so that when the convolution kernel slides, it only selects some positions in the holes for calculation, rather than all. This design enables the model to have a larger effective receptive field, that is, the convolution kernel perceives a wider range of input. Therefore, this method uses dilated convolutions with different dilation rates to extract features of targets of different scales, and combines the spatial pyramid structure to construct a multi-scale module MSFS (Multi-Scale Feature Selection).

[0056] The MSFS module adopts an innovative network structure to enhance the model's ability to perceive and recognize multi-scale features of images. The module consists of five parallel branches and dilated convolutions, which comprehensively extracts and utilizes features of different scales by cleverly fusing the output feature maps of each branch.

[0057] First, the one-dimensional convolution branch focuses on retaining the key features of the incoming image and providing basic information for subsequent processing. Then, the following three branches use dilated convolutions with different dilation rates (6, 12, 18). This convolution method has the characteristic of increasing the receptive field without introducing additional parameters, thereby capturing richer features at different scales.

[0058] Taking the convolution with a dilation rate of 6 as an example, if the input is a 3x3 convolution kernel, then when the dilated convolution is applied, two holes are actually inserted in the middle of the convolution kernel, making the final convolution kernel size 7x7, but the number of parameters remains at the original 9. In this way, the dilated convolution can effectively expand the receptive field of the convolution kernel without increasing the number of parameters, thereby capturing a wider range of contextual information.

[0059] At the same time, in order to further enhance the receptive field, the MSFS module introduces a hybrid dilated convolution. The output feature map of the dilated convolution with a dilated rate of 6 is input again into the dilated convolution with a dilated rate of 12 to maintain the close association between pixels, enhance the capture of image detail features, and reduce feature loss.

[0060] Similarly, the multi-scale representation of features is further optimized by inputting the output feature map of the dilated convolution with a dilation rate of 12 into the dilated convolution with a dilation rate of 18. This parallel method of dilated convolution enables the MSFS module to flexibly extract features at different scales by cleverly adjusting the hole size, thereby adapting to different target sizes and scene scales, and significantly improving the detection accuracy of multi-scale objects.

[0061] At the same time, by introducing the dilated convolution, the MSFS module achieves a larger receptive field while maintaining a small number of parameters, enhancing the contextual perception of the model and laying a solid foundation for improving overall performance. Finally, the maximum pooling branch provides the model with powerful overall image context insights by calculating the global semantic information of the entire feature map. This helps to improve the model's perception ability on the entire feature map, making it more context-aware.

[0062] By cleverly integrating the output feature maps of these five branches, the MSFS module achieves comprehensive utilization of features at different scales. This multi-scale feature extraction mechanism provides more comprehensive and accurate information for the detection model, significantly improving the performance of the model in multi-scale object detection tasks.

[0063] like Figure 3 As shown in the figure, facing the large number of multi-scale targets with different scales, the basic model YOLOv5s is prone to target loss when detecting multi-scale targets.

[0064] Based on this, the present application proposes a TFL (Target Feature Localization) location positioning module, which aims to enhance the target location features.

[0065] This module consists of two ConvBnSiLu ​​modules, a CA mixed domain attention mechanism, and a residual structure. The carefully designed structure optimizes the target positioning process and improves the accuracy of the detection model. The ConvBnSiLu ​​structure in the TFL module consists of a 2D convolution layer, a BN layer, and a SiLU activation function.

[0066] The first ConvBnSiLu ​​structure is used to map the input feature map to a more semantic representation space. Through the ConvBnSiLu ​​structure, the model can deeply understand the semantic content of the target feature and provide more targeted information for subsequent processing.

[0067] The second ConvBnSiLu ​​structure plays the role of dimensionality reduction. It performs a convolution operation on the feature map after the input feature map is fused with the feature map output by the CA mixed domain attention mechanism to match the previous feature map.

[0068] This step ensures that information from different sources can be effectively integrated, maintaining the model's comprehensive grasp of the target location characteristics. At the same time, the dimensionality reduction process helps reduce the computational burden and ensures effective integration of information.

[0069] The most important part of the TFL module is the CA mixed-domain attention mechanism. Its uniqueness lies in its comprehensive use of the channel-domain attention (Channel Attention) and spatial-domain attention (Spatial Attention) mechanisms.

[0070] This design aims to capture the features related to the target position more comprehensively and improve the accurate representation of the target position. First, the channel domain attention mechanism enables the model to understand the relationship between different channels. By calculating the importance of each channel and calculating the learned weights with the original feature map, the feature correlation between different channels is captured.

[0071] This means that the model can better understand the contribution of different feature channels to the target features, making the information of specific channels more prominent.

[0072] Second, the spatial domain attention mechanism focuses on assigning importance to different spatial locations in the feature map.

[0073] By globally pooling each spatial position and learning the weights, the model can pay more attention to the areas that have a greater impact on the target position. This attention mechanism helps capture the specific location information of the target in the image and reduce the detection error caused by position offset. Most importantly, the CA mixed domain attention mechanism introduces spatial information into the channel domain attention while integrating channel information.

[0074] This comprehensive design enables the module to comprehensively consider the channel and spatial relationship, so that the CA mixed domain attention mechanism can effectively solve the position and spatial characteristics of the detection target and optimize the model to detect the problem of unclear target position characteristics. It cooperates with the MSFS cross-scale perception module to improve the robustness and accuracy of the model.

[0075] Comparing this application with YOLOv5s, YOLOv5s+SPP, and YOLOv5s+SPPF, we can see that:

[0076] like Figure 4 As shown, in order to more comprehensively evaluate the advantages and disadvantages of the YOLOv5s+MSFS-TFL method proposed in this application compared with YOLOv5s, YOLOv5s+SPP, and YOLOv5s+SPPF in complex scenarios, the performance differences of these four methods in different categories of the data set will be compared based on the three categories of car, pedestrian, and cyclist on the KITTI data set.

[0077] The YOLOv5s model performs 97.10 on the Car category, 93.30 on the Pedestrian category, and 81.70 on the Cyclist category.

[0078] Compared with the basic model, YOLOv5s+SPP has achieved slight improvements in all categories, especially in the Cyclist category, where the detection accuracy has increased to 82.70.

[0079] Compared with the previous two models, the detection accuracy of YOLOv5s+SPPF model in all categories remains at a similar level, which are 97.20, 92.80, and 82.30 respectively.

[0080] Compared with YOLOv5s, YOLOv5s+SPP and YOLOv5s+SPPF, the performance changes are small and there is no obvious improvement. The YOLOv5s+MSFS-TFL model achieves the highest detection accuracy in all categories, especially in the Pedestrian category, reaching 93.30.

[0081] Compared with other models, YOLOv5s+MSFS-TFL performs better in all categories.

[0082] like Figure 5 As shown in the figure, the experimental results show that the YOLOv5s model combined with the MSFS-TFL module has achieved significant improvements in multiple performance indicators, especially in the average precision (mAP50) and precision (Precision) with an intersection-over-union ratio of 0.5, which is significantly better than the traditional multi-scale module SPP and its variant SPPF fusion strategy.

[0083] Specifically, when the MSFS-TFL module is introduced, the model reaches 91.80 on mAP50, which is 1.1%, 0.8%, and 1.1% higher than YOLOv5s alone and YOLOv5s combined with SPP and SPPF modules, respectively.

[0084] This result highlights the significant role of the MSFS-T module in optimizing model performance, especially its efficiency in dealing with multi-scale object detection problems.

[0085] In terms of the Recall indicator, YOLOv5s-MSFS-TFL also showed relatively high performance. Although the difference was not significant compared with YOLOv5s-SPP, combined with the improvement of other performance indicators, it fully demonstrated the overall advantages of the MSFS-T module.

[0086] More importantly, in terms of Precision, YOLOv5s-MSFS-TFL reached 93.80, significantly surpassing other configurations. This result not only verifies the effectiveness of the MSFS-T module in improving detection accuracy, but also reflects its potential in reducing false positive rates and optimizing detection results.

[0087] like Figure 6 As shown, in order to intuitively compare the detection performance differences between the multi-scale target detection model YOLOv5s-MSFS-T and the original model YOLOv5s, this application performed a visual analysis on the KITTI dataset.

[0088] The original YOLOv5s model is shown on the left, and the multi-scale detection model YOLOv5s-MSFS-T proposed in this application is shown on the right. By comparing the multi-scale detection results of the two images, this application observes that the YOLOv5s-MSFS-T model is significantly better than the original YOLOv5s model on the KITTI dataset.

[0089] In the right image, the YOLOv5s-MSFS-T model successfully detects the target that was not recognized in the left image, demonstrating excellent target detection performance.

[0090] In addition, the detection accuracy of the right image is relatively improved. This application adds the MSFS-T module to the Neck part of the YOLOv5s model. The introduction of the cross-scale perception module MSFS gives the model stronger adaptability, especially for the detection of small-sized targets. The integration of the position positioning module TFL further enhances the positioning accuracy of the target, making the model perform well in target detection tasks.

[0091] In summary, by comparing the results of the two images, it is clear that the YOLOv5s-MSFS-T model has significant advantages in multi-scale object detection.

[0092] Obviously, those skilled in the art can make various modifications and changes to the embodiments of the present application without departing from the spirit and scope of the present application. In this way, if these modifications and changes are within the scope of the claims of the present application and their equivalents, the present application is also intended to cover these modifications and changes. The word "comprising" does not exclude the presence of other elements or steps not listed in the claims. The simple fact that certain measures are recorded in mutually different dependent claims does not indicate that the combination of these measures cannot be used to profit. Any figure mark in the claims should not be considered to limit the scope.

Claims

1. A multi-scale target detection method based on YOLOv5s, characterized in that: The following steps are involved: S1, dataset acquisition; S2, preprocessing the data set; S3, build a MSFS module with cross-scale perception; S4, constructing a TFL module for location positioning; S5, embedding the MSFS module and the TFL module into Neck to obtain an optimized multi-scale target detection model based on YOLOv5s; S6. Input the preprocessed data set into the optimized multi-scale object detection model based on YOLOv5s to obtain the final data set.

2. A multi-scale target detection method based on YOLOv5s according to claim 1, characterized in that, The MSFS module uses dilated convolutions with different dilation rates to extract features of targets of different scales, and constructs the module MSFS in combination with a spatial pyramid structure.

3. A multi-scale target detection method based on YOLOv5s according to claim 2, characterized in that, The MSFS module consists of five parallel branches and dilated convolutions with different dilation rates.

4. A multi-scale target detection method based on YOLOv5s according to claim 3, characterized in that: The five parallel branches in the MSFS module are used to input the received preprocessed data set into one-dimensional convolution, dilated convolution with a dilated rate of 6, dilated convolution with a dilated rate of 8, dilated convolution with a dilated rate of 12, and maximum pooling, respectively.

5. A multi-scale target detection method based on YOLOv5s according to claim 4, characterized in that: The S3 specific method is as follows: The preprocessed data sets are input into five parallel branches respectively; the preprocessed data sets input into the first branch enter into a one-dimensional convolution to save the key features in the preprocessed data sets; The preprocessed data sets input into the second to fifth branches enter the first dilated convolution, the second dilated convolution, the third dilated convolution with the dilated rate, and the maximum pooling respectively.

6. A multi-scale target detection method based on YOLOv5s according to claim 5, characterized in that: The S3 specific method also includes: the preprocessed data set is subjected to a first dilated convolution to output a first feature map, the first feature map is input into a second dilated convolution to output a second feature map, and the second feature map is input into a third dilated convolution to output a third feature map.

7. A multi-scale target detection method based on YOLOv5s according to claim 1, characterized in that: The TFL module consists of two ConvBnSiLU modules, a CA mixed-domain attention mechanism and a residual structure.

8. A multi-scale target detection method based on YOLOv5s according to claim 7, characterized in that: The ConvBnSiLU module consists of a 2D convolutional layer, a BN layer and a SiLU activation function.

9. A multi-scale target detection method based on YOLOv5s according to claim 8, characterized in that: The specific method of S4 is as follows: The feature map input into the TFL module is mapped to a more semantically rich representation space through the first ConvBnSiLU module. The mapped feature map is output to the second ConvBnSiLU module through the CA mixed domain attention mechanism. The second ConvBnSiLU module performs a convolution operation on the feature map obtained by fusing the input feature map and the feature map output by the CA mixed domain attention mechanism.

10. A multi-scale target detection method based on YOLOv5s according to claim 9, characterized in that: The CA mixed domain attention mechanism is used to assign weights of different spatial positions to the input feature map.

Images