Small-size vehicle detection deep learning model based on multi-scale module feature fusion

Abstract

The invention provides a small-size vehicle detection deep learning model based on multi-scale module feature fusion, which solves the problems of small-size vehicle image detection and the like, and comprises a Backbone network, a Neck layer and a Head network, a C2fDCNv3 module based on the combination of a variable convolution DCNv3 and a double-convolution cross-stage module C2f and an SPPFLSKA module based on the combination of a spatial pyramid pooling layer SPPF and a large separable convolution kernel attention module LSKA are introduced into the Backbone network. A C2fSCConv module based on the combination of a space channel reconstruction convolution SCConv and a double convolution cross-stage module C2f is introduced into the Neck layer; and a multi-scale convolution kernel detection module MSKDeect is introduced into the Head network. The method has the advantages of good small-size image detection effect, cost reduction and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle recognition and detection, and specifically relates to a small-size vehicle detection deep learning model based on multi-scale module feature fusion. Background Art

[0002] As a key branch in the field of computer vision, object detection algorithms have undergone significant technological innovation and progress in recent years. In the early days, the field of object detection mainly adopted a two-stage processing method, with R-CNN and Faster-R-CNN as typical representatives. This method selects candidate regions in the image, then extracts features and classifies each region separately. Although this method is effective, it has a slow processing speed. Subsequently, in order to solve the speed problem, single-stage object detection algorithms such as the famous YOLO series and SSD emerged. This method significantly improves the processing speed by completing the entire detection task in one forward propagation of the network. It regards object detection as a regression problem, directly predicting bounding boxes and class probabilities in the image, thereby achieving fast and efficient detection. The current more advanced YOLOv8 model has demonstrated excellent results and performance in the target detection task of conventional vehicles in the transportation field. However, when faced with small-sized vehicle targets, its detection effect is poor, and it is impossible to quickly and accurately identify and locate them.

[0003] In order to solve the shortcomings of the existing technology, people have conducted long-term exploration and proposed various solutions. For example, a Chinese patent document discloses a drone small target image detection method based on MultiHeads-YOLOv5 [202310889055.1], which is implemented by the following steps: Step 1, build the same FPN+PAN up- and down-sampling structures, which include common structures of deep learning networks such as Conv, Upsample, and Concat, as well as C3 modules and Conv, Concat, and C3 up-sampling and down-sampling network structures. These structures are used to extract the original model feature form. Step 2, for the Neck architecture of the original model YOLOv5, refer to the FPN+PAN structure built in step 1, deepen the network structure depth of the original model, so that more subtle feature information can be extracted for detecting small target scale objects. Step 3, through the C3 module of the newly added up- and down-sampling structure in the Neck part of the model, connect a detection head in YOLOv5 to complete the detection task of micro-scale small targets.

[0004] The above solution solves the problem of target detection to a certain extent, but the solution still has many shortcomings, such as the inability to quickly identify and locate small-sized vehicles. Summary of the invention

[0005] The purpose of the present invention is to provide a small-size vehicle detection deep learning model based on multi-scale module feature fusion, which has a reasonable design and good small-size vehicle recognition and positioning effect in order to solve the above problems.

[0006] To achieve the above-mentioned purpose, the present invention adopts the following technical scheme: a deep learning model for small-size vehicle detection based on multi-scale module feature fusion, including a Backbone network, a Neck layer and a Head network. The Backbone network introduces a C2f_DCNv3 module based on the combination of variable convolution DCNv3 and a double convolution cross-stage module C2f, and an SPPF_LSKA module based on the combination of a spatial pyramid pooling layer SPPF and a large separable convolution kernel attention module LSKA to extract features of small-size vehicle images; the Neck layer introduces a C2f_SCConv module based on the combination of spatial channel reconstruction convolution SCConv and a double convolution cross-stage module C2f to process the features extracted by the Backbone network; the Head network introduces a multi-scale convolution kernel detection module MSK_Detect to predict the extracted small-size vehicle image features.

[0007] In the above-mentioned small-size vehicle detection deep learning model based on multi-scale module feature fusion, the input and output of the C2f_DCNv3 module pass through the convolution layer Conv, and the bottleneck layer in the middle of the C2f_DCNv3 module introduces 2 layers of deformable convolution DCNv3.

[0008] In the above-mentioned small-size vehicle detection deep learning model based on multi-scale module feature fusion, the specific calculation process of deformable convolution DCNv3 is as follows:

[0009]

[0010] Where y0 represents the output result of the DCNv3 operation on the value of the feature map at position p0; G represents the number of aggregation groups, which means that DCNv3 uses multiple spatial positions to learn multiple spatial aggregation modes; K represents the total number of sampling areas; W g Represents the convolution kernel weight in the sampling operation; M gk ∈(0, 1) represents the modulation scalar, which is a normalized coefficient used to adjust the contribution weight of each sampling area; x g represents the input feature map; p k Indicates the position of the kth sampling area in the convolution kernel; Δp gk Represents the offset of the sampling area, allowing the convolution kernel to adaptively adjust its shape.

[0011] In the above-mentioned small-size vehicle detection deep learning model based on multi-scale module feature fusion, the input and output of the C2f_SCConv module pass through the convolution layer Conv. The bottleneck layer in the middle of the C2f_SCConv module introduces two layers of space and channel reconstruction convolution SCConv. The space and channel reconstruction convolution SCConv is composed of a space reconstruction module SRU and a channel reconstruction module CRU.

[0012] In the above-mentioned small-size vehicle detection deep learning model based on multi-scale module feature fusion, the formula of GroundNorm in the spatial reconstruction module SRU is:

[0013]

[0014] where γ∈R C It is a trainable affine transformation. The information weight W∈R is obtained by normalizing γ to Standard. C , perform a threshold-based Gate operation on W to obtain the information-rich weight W up With less information weight W low The input is passed through these weights to obtain X1 and X2, which are split into 1 / 2 based on the channel, and then interleaved and multiplied respectively and concatenated based on the channel.

[0015] In the above-mentioned small-size vehicle detection deep learning model based on multi-scale module feature fusion, the channel reconstruction module CRU performs α-based segmentation on the input data, and then performs upper-layer convolution operations and lower-layer convolution operations respectively. The upper-layer convolution operation includes k×k and 1×1 convolution kernels, and the lower-layer operation includes 1×1 convolution kernels and simple skip-connection. After subsequent operations based on AvgPooling and Softmax, channel splicing is performed.

[0016] In the above-mentioned small-size vehicle detection deep learning model based on multi-scale module feature fusion, after the SPPF_LSKA module passes through the first convolution layer Conv, it passes through 3 MaxPooling layers, splices the output obtained by each MaxPooling layer, and then passes through the large separable convolution kernel attention module LSKA and the convolution layer Conv to obtain the output.

[0017] In the above-mentioned small-size vehicle detection deep learning model based on multi-scale module feature fusion, the large separable convolution kernel attention module LSKA is constructed by multiple convolution layers Conv. The convolution kernels in each convolution layer Conv are different, and they are horizontal convolution kernel 1×(2d-1), vertical convolution kernel (2d-1)×1, horizontal convolution kernel 1×k / d, vertical convolution kernel k / d×1, in order, where k represents the original convolution kernel size, d represents the convolution void ratio; the last one is a 1×1 unit convolution kernel.

[0018] In the above-mentioned small-size vehicle detection deep learning model based on multi-scale module feature fusion, the multi-scale convolution kernel detection module MSK_Detect performs channel separation operation through the split operation. The output obtained by split will pass through the convolution layer Conv of multiple convolution kernels of different sizes; after decoupling through the convolution layer Conv of the 1×1 convolution kernel, two 1×1 convolution layers Conv are used to obtain the output results of the target rectangular box and the target classification.

[0019] In the above-mentioned small-size vehicle detection deep learning model based on multi-scale module feature fusion, this small target detection deep learning model is improved based on the target detection model yolov8.

[0020] Compared with the existing technology, the advantages of the present invention are: in the Backbone network, the C2f_DCNv3 module is introduced to realize rich feature extraction capabilities, which provides a powerful and efficient Backbone network foundation for the whole model, ensuring the high performance and robustness of the overall model when processing small-size vehicle image feature extraction tasks; in the Backbone network, the SPPF_LSKA module is introduced into the deep structure of the Backbone network. While fusing deep receptive fields of different scales, the SPPF_LSKA module captures more long-range dependencies in small-size vehicle feature maps based on the large separable convolution kernel attention module LSKA, as well as better spatial and channel adaptation. Adaptability; The combination of block construction has the advantages of high efficiency, easy integration, optimized parameter configuration and reduced parameter amount, which reduces the computing and storage costs; In the Neck layer, the C2f_SCConv module is introduced. The C2f_SCConv module has the advantages of high efficiency, easy integration, optimized parameter configuration and reduced parameter amount. While reducing the computing and storage costs, it provides better performance than the traditional convolution operation; In the Head network, the multi-scale convolution kernel detection module MSK_Detect is introduced to improve the diversity of small-size vehicle image feature expression and enhance the network's generalization ability in the face of complex features. At the same time, the design of the multi-scale convolution kernel detection module MSK_Detect reduces the number of parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is the overall architecture diagram of the model of the present invention.

[0022] Figure 2 It is a structural schematic diagram of the C2f_DCNv3 module and the variable convolution DCNv3 of the present invention.

[0023] Figure 3 It is a structural schematic diagram of the C2f_SCConv module and the spatial channel reconstruction convolution SCConv of the present invention.

[0024] Figure 4 It is a structural schematic diagram of the SPPF_LSKA module of the present invention and the large separable convolution kernel attention module LSKA.

[0025] Figure 5 It is a structural schematic diagram of the multi-scale convolution kernel detection module MSK_Detect of the present invention.

[0026] Figure 6 It is a comparative schematic diagram of small-sized vehicle detection of the present invention. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] like Figure 1 As shown in the figure, a deep learning model for small-sized vehicle detection based on multi-scale module feature fusion can improve the safety of autonomous driving and help vehicles quickly and accurately identify small obstacles such as road signs, pedestrians, and bicycles, thereby better ensuring driving safety. On the other hand, it can also improve the efficiency of security monitoring and help the monitoring system quickly and accurately identify small targets such as suspicious persons and vehicles, thereby better ensuring public safety. It specifically includes Backbone network, Neck layer and Head network. The Backbone network introduces the C2f_DCNv3 module based on the combination of variable convolution DCNv3 and double convolution cross-stage module C2f and the SPPF_LSKA module based on the combination of spatial pyramid pooling layer SPPF and large separable convolution kernel attention module LSKA to extract features of small-sized vehicle images; the Neck layer introduces the C2f_SCConv module based on the combination of spatial channel reconstruction convolution SCConv and double convolution cross-stage module C2f to process the features extracted by the Backbone network; the Head network introduces the multi-scale convolution kernel detection module MSK_Detect to predict the extracted small-sized vehicle image features. The above-mentioned small target detection deep learning model can effectively handle the detection task of small-sized vehicle targets in complex images.

[0029] like Figure 2As shown in the figure, the structure of the C2f_DCNv3 module based on the combination of the variable convolution DCNv3 and the dual convolution cross-stage module C2f, as well as the overall structure of the variable convolution DCNv3 are respectively represented. Figure 2 The left part represents the propagation structure of the C2f_DCNv3 module. First, the input and output of the C2f_DCNv3 module pass through the convolution layer Conv. The bottleneck layer in the middle of the C2f_DCNv3 module introduces 2 layers of deformable convolution DCNv3, where the bottleneck layer can introduce various types of multi-layer modules. Figure 2 The right part describes the structure of the variable convolution DCNv3. Based on the tensor operation propagation operation of the offset weight offset and the scalar weight mask, the sampling graph sampling_grid is obtained as the sampling operation area performed by the variable convolution DCNv3, and finally the output of the variable convolution DCNv3 is obtained.

[0030] In depth, the specific calculation process of deformable convolution DCNv3 is as follows:

[0031]

[0032] Where y0 represents the output result of the DCNv3 operation on the value of the feature map at position p0; G represents the number of aggregation groups, which means that DCNv3 uses multiple spatial positions to learn multiple spatial aggregation modes; K represents the total number of sampling areas; W g Represents the convolution kernel weight in the sampling operation; represents the modulation scalar, which is a normalized coefficient used to adjust the contribution weight of each sampling area; x g represents the input feature map; p k Indicates the position of the kth sampling area in the convolution kernel; Δp gk Represents the offset of the sampling area, allowing the convolution kernel to adaptively adjust its shape.

[0033] Based on the constructed C2f_DCNv3 module, as the key part of the Backbone network of the model of the present invention, it has a powerful feature extraction capability. The module effectively processes large-scale image data by introducing long-distance dependence and adaptive spatial aggregation mechanism, and then learns a more powerful and robust feature representation. The C2f_DCNv3 module has higher parameter usage efficiency. Through the introduction of weight sharing and multi-group mechanisms, it realizes rich feature extraction capabilities while ensuring parameter and computational efficiency. In addition, the C2f_DCNv3 module shows excellent stability during the training process, effectively reducing the unstable factors that may occur during training. This improvement in stability not only optimizes the training process, but also helps to improve the generalization ability and reliability of the model. In summary, the addition of the C2f_DCNv3 module provides a powerful and efficient Backbone network foundation for the model of the present invention, ensuring the high performance and robustness of the overall model when processing complex visual tasks.

[0034] like Figure 3 As shown in the figure, the structure of the C2f_SCConv module constructed based on the combination of the spatial channel reconstruction convolution SCConv and the double convolution cross-stage module C2f, as well as the overall structure of the spatial channel reconstruction convolution SCConv module are respectively represented. Figure 3 The left part of the figure shows the propagation structure of the C2f_SCConv module, which is similar to the framework of C2f in the C2f_DCNv3 module. The input and output of the C2f_SCConv module pass through the convolution layer Conv. The bottleneck layer in the middle of the C2f_SCConv module introduces two layers of space and channel reconstruction convolution SCConv. The space and channel reconstruction convolution SCConv is composed of the space reconstruction module SRU and the channel reconstruction module CRU.

[0035] Furthermore, Figure 3 The middle part represents the structure of the spatial reconstruction module SRU. The formula of GroundNorm in the spatial reconstruction module SRU is:

[0036]

[0037] where γ∈R C It is a trainable affine transformation. The information weight W∈R is obtained by normalizing γ to Standard. C , perform a threshold-based Gae operation on W to obtain the information-rich weight W up With less information weight W low The input is passed through these weights to obtain X1 and X2, which are split into 1 / 2 based on the channel, and then interleaved and multiplied respectively and concatenated based on the channel.

[0038] In addition, Figure 3The right part of the figure represents the structure of the channel reconstruction module CRU. After the channel reconstruction module CRU performs α-based segmentation on the input data, it performs upper-layer convolution operations and lower-layer convolution operations respectively. The upper-layer convolution operation includes k×k and 1×1 convolution kernels, and the lower-layer operation includes 1×1 convolution kernels and simple skip-connection. After subsequent operations based on AvgPooling and Softmax, channel splicing is performed.

[0039] The spatial channel reconstruction convolution SCConv enhances the information flow between features through separation and reconstruction methods, as well as cross reconstruction methods, thereby improving the efficiency and expression ability of feature extraction, and can reduce spatial redundancy. The spatial channel reconstruction convolution SCConv can more accurately capture and express important information in the image, especially when processing complex and subtle features. In the present invention, the C2f_SCConv module is introduced in the Neck layer. The C2f_SCConv module has the advantages of high efficiency, easy integration, optimized parameter configuration and reduced parameter quantity. While reducing computing and storage costs, it provides better performance than traditional convolution operations.

[0040] As attached Figure 4 As shown in the figure, the overall structure of the SPPF_LSKA module based on the large separable convolution kernel attention LSKA and the spatial pyramid pooling layer SPPF, as well as the internal structure of the large separable convolution kernel attention LSKA, are respectively represented. Figure 4 The left part of the figure shows the propagation path of the SPPF_LSKA module structure. After the SPPF_LSKA module passes through the first convolution layer Conv, it passes through three MaxPooling layers, concatenates the outputs of each MaxPooling layer, and then passes through the large separable convolution kernel attention module LSKA and the convolution layer Conv to obtain the output.

[0041] Visibly, attached Figure 4 The right part of the figure shows the internal structure of the large separable convolution kernel attention LSKA. The large separable convolution kernel attention module LSKA is constructed by multiple convolution layers Conv. The convolution kernels in each convolution layer Conv are different. In order, they are horizontal convolution kernel 1×(2d-1), vertical convolution kernel (2d-1)×1, horizontal convolution kernel 1×k / d, vertical convolution kernel k / d×1, where k represents the original convolution kernel size and d represents the void ratio of the convolution; the last one is a 1×1 unit convolution kernel.

[0042] LSKA aims to solve the problem that the computation and memory usage of deep convolutional layers increase linearly with the size of the convolution kernel. LSKA decomposes the 2D convolution kernel into horizontal and vertical 1D convolution kernels, which enables the direct use of deep convolution layers with large convolution kernels without any additional modules. LSKA decomposes the 2D weight kernels in deep convolution and deep dilated convolution into two cascaded 1D separable weight kernels. It not only reduces the computational complexity, but also maintains performance comparable to traditional large convolution kernel modules.

[0043] The advantages of large separable convolution kernel attention LSKA are: 1. Long-range dependency: LSKA effectively captures long-range dependencies in images, which is a key component implemented by the self-attention mechanism in the transformer. LSKA provides an alternative method to capture global information through large convolution kernel design; 2. Spatial and channel adaptability: LSKA inherits the spatial attention and channel attention strategies of traditional large convolution kernels. These two strategies adaptively recalibrate the weights of features according to contextual dependencies. Compared with self-attention and LKA modules, LSKA has better spatial and channel adaptability while keeping parameters and computational complexity low; 3. Reduced computational complexity and memory usage: Compared with standard traditional large convolution kernel modules, the computational complexity and memory usage are significantly reduced.

[0044] In the present invention, the SPPF_LSKA module is introduced into the deep structure of the Backbone network. While fusing deep receptive fields of different scales, more long-range dependencies in features are captured based on the LSKA module, as well as better spatial and channel adaptability.

[0045] As attached Figure 5 As shown, the structure of the multi-scale convolution kernel detection head MSK_Detect is shown. The multi-scale convolution kernel detection module MSK_Detect performs channel separation operation through split operation. The output obtained by split will pass through the convolution layer Conv of multiple convolution kernels of different sizes. In the present invention, two types of convolution kernels, 3×3 and 5×5, will be selected to form a multi-scale convolution kernel; after decoupling through the convolution layer Conv of the 1×1 convolution kernel, two 1×1 convolution layers Conv are used to obtain the output results of the target rectangular frame and the target classification.

[0046] The multi-scale convolution kernel of the MSK module enables the capture of features at different levels, thereby improving the diversity of feature expression, enhancing the generalization ability of the network in the face of complex features, and enabling the model to exhibit better performance and adaptability in a wider range of application scenarios; at the same time, the design of the MSK module reduces the number of module parameters. By optimizing the use and fusion strategies of convolution kernels, the MSK module optimizes parameter calculation efficiency, thereby reducing computational costs and space occupancy. In the present invention, the MSK module is introduced into the head layer to construct a multi-scale convolution kernel detection module MSK_Detect. By applying multi-scale convolution kernels, the MSK_Detect module can effectively improve the ability to capture and process features of different sizes regarding small-sized vehicle targets. This multi-scale approach enables the module to understand the content of small-sized vehicles at different levels, thereby improving the accuracy and robustness of small-sized vehicle detection.

[0047] Preferably, this small target detection deep learning model is improved based on the target detection model yolov8 to verify the effectiveness of the method, and is tested through the following comparative experiments.

[0048] The deep learning experimental process is constructed based on the large-scale real-life scene image dataset visdrone, which contains small-sized vehicles. The samples in this dataset are diverse and abundant, and include multiple types of vehicle categories. The yolov8 algorithm and the proposed small-scale target detection model based on multi-scale module feature fusion are trained and tested based on the test set. The experimental test results show that on the test dataset, based on the mean Average Precision (mAP) evaluation index, the yolov8 model achieved 49.7% mAP50 and 30.3% overall mAP, while the model proposed in this application achieved 52.5% mAP and 32.1% overall mAP. These results show that the model proposed in this application shows better performance than the traditional yolov8 method in the detection task of small-sized vehicle targets in large-scale images of real scenes.

[0049] The following table shows the performance comparison of the YOLOv8 model and the small-size vehicle detection deep learning model based on multi-scale module feature fusion proposed in this application on a small-size vehicle target detection dataset, indicating the improvement of the method proposed in this application relative to performance.

[0050] Precision Recall mAP50 mAP50-95 YOLOv8l 60.74％ 49.77％ 49.71％ 30.32％ YOLOv8x 61.01％ 50.24％ 50.32％ 31.03％ Ours 62.62％ 52.85％ 52.51％ 32.12％

[0051] Figure 6This is an example of the comparison of small-sized vehicle target detection results in real-life scenarios. Part (a) of the figure shows the detection results of the original yolov8 model, and part (b) shows the detection results of the proposed new model. The model of this application can better detect small-sized vehicle targets than the yolov8 model.

[0052] To sum up, the principle of this embodiment is: based on the overall structural style of YOLOv8, improvements in multiple aspects are made to construct a deep network feature detection head for the size of small target vehicles, so that the model is more suitable for learning the characteristic size of small target vehicles; deformable convolution DCNv3 is introduced in the backbone network to improve the processing capability and stability of input features; large separable convolution kernel attention module LSKA is introduced in the deep stage of the backbone network to better capture the global information of features; spatial channel reconstruction convolution SCConv is introduced in the Neck layer, which can effectively reduce the redundancy of the convolutional network while improving the performance and efficiency of the network, and has a better recognition and positioning effect for small-sized vehicle targets.

[0053] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0054] Although this article uses the terms Backbone network, Neck layer and Head network more frequently, it does not exclude the possibility of using other terms. These terms are used only to more conveniently describe and explain the essence of the present invention; interpreting them as any additional restrictions is contrary to the spirit of the present invention.

Claims

1. A small-size vehicle detection deep learning model based on multi-scale module feature fusion, including a Backbone network, a Neck layer and a Head network, characterized in that: The Backbone network introduces the C2f_DCNv3 module based on the combination of the variable convolution DCNv3 and the double convolution cross-stage module C2f, and the SPPF_LSKA module based on the combination of the spatial pyramid pooling layer SPPF and the large separable convolution kernel attention module LSKA to extract features of small-sized vehicle images; the Neck layer introduces the C2f_SCConv module based on the combination of the spatial channel reconstruction convolution SCConv and the double convolution cross-stage module C2f to process the features extracted by the Backbone network; the Head network introduces the multi-scale convolution kernel detection module MSK_Detect to predict the extracted small-sized vehicle image features.

2. The small-size vehicle detection deep learning model based on multi-scale module feature fusion according to claim 1 is characterized in that: The input and output of the C2f_DCNv3 module pass through the convolution layer Conv, and the bottleneck layer in the middle of the C2f_DCNv3 module introduces 2 layers of deformable convolution DCNv3.

3. The small-size vehicle detection deep learning model based on multi-scale module feature fusion according to claim 2 is characterized in that: The specific calculation process of the deformable convolution DCNv3 is as follows: Where y0 represents the output result of the DCNv3 operation on the value of the feature map at position p0; G represents the number of aggregation groups, which means that DCNv3 uses multiple spatial positions to learn multiple spatial aggregation modes; K represents the total number of sampling areas; W g Represents the convolution kernel weight in the sampling operation; M gk ∈(0, 1) represents the modulation scalar, which is a normalized coefficient used to adjust the contribution weight of each sampling area; x g represents the input feature map; p k Indicates the position of the kth sampling area in the convolution kernel; Δp gk Represents the offset of the sampling area, allowing the convolution kernel to adaptively adjust its shape.

4. The small-size vehicle detection deep learning model based on multi-scale module feature fusion according to claim 1 is characterized in that: The input and output of the C2f_SCConv module pass through the convolution layer Conv. The bottleneck layer in the middle of the C2f_SCConv module introduces two layers of space and channel reconstruction convolution SCConv. The space and channel reconstruction convolution SCConv is composed of a space reconstruction module SRU and a channel reconstruction module CRU.

5. The small-size vehicle detection deep learning model based on multi-scale module feature fusion according to claim 4 is characterized in that: The formula of GroundNorm in the spatial reconstruction module SRU is: where γ∈R C It is a trainable affine transformation. The information weight W∈R is obtained by normalizing γ to Standard. C , perform a threshold-based Gate operation on W to obtain the information-rich weight W up With less information weight W low The input is passed through these weights to obtain X1 and X2, which are split into 1 / 2 based on the channel, and then interleaved and multiplied respectively and concatenated based on the channel.

6. The small-size vehicle detection deep learning model based on multi-scale module feature fusion according to claim 4 is characterized in that: After the channel reconstruction module CRU performs α-based segmentation on the input data, it performs upper-layer convolution operations and lower-layer convolution operations respectively, wherein the upper-layer convolution operation includes k×k and 1×1 convolution kernels, and the lower-layer operation includes 1×1 convolution kernels and simple skip-connection. After subsequent AvgPooling and Softmax operations, channel splicing is performed.

7. The small-size vehicle detection deep learning model based on multi-scale module feature fusion according to claim 1 is characterized in that: The SPPF_LSKA module passes through the first convolution layer Conv and then three MaxPooling layers, concatenating the outputs of each MaxPooling layer, and then passes through the large separable convolution kernel attention module LSKA and the convolution layer Conv to obtain the output.

8. The small-size vehicle detection deep learning model based on multi-scale module feature fusion according to claim 7 is characterized in that: The large separable convolution kernel attention module LSKA is constructed by multiple convolution layers Conv, and the convolution kernels in each convolution layer Conv are different, which are horizontal convolution kernel 1×(2d-1), vertical convolution kernel (2d-1)×1, horizontal convolution kernel 1×k / d, vertical convolution kernel k / d×1, in order, where k represents the original convolution kernel size, d represents the convolution void ratio; the last one is a 1×1 unit convolution kernel.

9. The small-size vehicle detection deep learning model based on multi-scale module feature fusion according to claim 1 is characterized in that: The multi-scale convolution kernel detection module MSK_Detect performs a channel separation operation through a split operation. The output obtained by split will pass through a convolution layer Conv of multiple convolution kernels of different sizes; after decoupling through a convolution layer Conv of a 1×1 convolution kernel, two 1×1 convolution layers Conv are used to obtain the output results of the target rectangular box and the target classification.

10. The small-size vehicle detection deep learning model based on multi-scale module feature fusion according to claim 1 is characterized in that: This small target detection deep learning model is improved based on the target detection model yolov8.

Images