Small-sized vehicle detection deep learning model based on fusion of multi-scale module characteristics

Abstract

To provide a small-sized vehicle detection deep learning model based on fusion of multi-scale module characteristics, which detects images of small-sized vehicles.SOLUTION: An architecture includes a backbone network, a neck layer, and a head network, where a C2f_DCNv3 module, which is based on a combination of a deformable convolutional DCNv3 and a double convolutional cross-stage module C2f, and an SPPF_LSKA module, which is based on a combination of a spatial pyramid pooling layer SPPF and a large separable convolutional kernel attention module LSKA, is introduced into the backbone network, a C2f_SCConv module, which is based on a combination of the double convolutional cross-stage module C2f and a spatial channel reconstruction convolutional SCConv, is introduced into the neck layer, and a multi-scale convolutional kernel detection module MSK_Detect is introduced into the head network.SELECTED DRAWING: Figure 1

Description

[Technical Field]

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

[0002] As an important area in the field of computer vision, object detection algorithms have undergone significant technological innovation and advancements in recent years. Early object detection techniques primarily employed two-stage processing methods, such as Region-Convolutional Neural Network (R-CNN) and Faster-R-CNN. These methods select candidate regions within an image and then extract and classify features for each region. While effective, these methods suffer from slow processing speeds. To address this speed issue, single-stage object detection algorithms, such as the well-known You Only Look Once (YOLO) family and the Single Shot Multibox Detector (SSD), emerged. These methods significantly improve processing speed by completing the entire detection task in a single forward pass of the network. Object detection is treated as a regression problem, and bounding boxes and category probabilities are predicted directly within the image, achieving fast and efficient detection. The relatively advanced YOLOv8 model has demonstrated excellent results and performance in general vehicle object detection tasks in the transportation domain. However, when faced with small vehicle objects, the detection effect is poor, making it difficult to quickly and accurately identify and locate them.

[0003] To address the shortcomings of the prior art, researchers have conducted long-term research and proposed various solutions. For example, Patent Document 1 discloses a drone-based small object image detection method based on MultiHeads-YOLOv5, which is realized in the following steps: In step 1, the same FPN+PAN upsampling and downsampling structure is constructed. These structures include common deep learning network structures such as Conv, Upsample, and Concat, a C3 module, and two types of upsampling and downsampling network structures, Conv, Concat, and C3. These structures are used to extract feature shapes from the original model. In step 2, the FPN+PAN structure constructed in step 1 is used in the neck architecture of the original model, YOLOv5, to deepen the depth of the original model's network structure, thereby extracting smaller feature information and detecting small-scale objects. In step 3, the YOLOv5 detection head is connected through the C3 module in the new upsampling and downsampling structure added to the neck of the model to complete the task of detecting small micro-scale objects.

[0004] Although the above technical solutions have solved the problem of object detection to some extent, the technical solutions still have many shortcomings, such as the inability to quickly identify and locate small vehicles. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Chinese Patent Application Publication No. 117011727 Summary of the Invention [Problem to be solved by the invention]

[0006] In response to the above problems, the objective of the present invention is to provide a rationally designed deep learning model for small vehicle detection based on the fusion of multi-scale modular features, which has excellent recognition and localization effects for small vehicles. [Means for solving the problem]

[0007] To achieve the above objectives, the present invention employs the following technical proposal: a small vehicle detection deep learning model based on multi-scale module feature fusion, which includes a backbone network, a neck layer, and a head network, in which the backbone network introduces a C2f_DCNv3 module based on a combination of a deformable convolution DCNv3 and a double convolution cross-stage module C2f, and an SPPF_LSKA module based on a combination of a spatial pyramid pooling layer SPPF and a large separable convolution kernel attention module LSKA to extract features from images of small vehicles, the neck layer introduces a C2f_SCConv module based on a combination of a spatial channel reconstruction convolution SCConv and a double convolution cross-stage module C2f to process the features extracted by the backbone network, and the head network introduces a multi-scale convolution kernel detection module MSK_Detect to predict the extracted image features of small vehicles.

[0008] In the small vehicle detection deep learning model that combines the features of the above multi-scale modules, the input and output of the C2f_DCNv3 module pass through the convolutional layer Conv, and a two-layer deformable convolutional DCNv3 is introduced in the middle bottleneck layer of the C2f_DCNv3 module.

[0009] In the above-mentioned small vehicle detection deep learning model based on multi-scale modular feature fusion, the specific calculation process of deformable convolution DCNv3 is as follows:

[0010]

number

[0011] where: JPEG2025181612000003.jpg66 is the feature map JPEG2025181612000004.jpg56 indicates the output result calculated by DCNv3 calculation. JPEG2025181612000005.jpg66 represents the number of aggregation groups, and indicates that multiple spatial locations are used to learn multiple spatial aggregation modes in DCNv3 operation. JPEG2025181612000006.jpg64 represents the total number of sampling areas, JPEG2025181612000007.jpg66 represents the weight of the convolution kernel in the sampling operation, JPEG2025181612000008.jpg621 represents the modulation scalar, which is the normalized coefficient for adjusting the contribution weight of each sampling region, JPEG2025181612000009.jpg77 represents the input feature map, JPEG2025181612000010.jpg76 represents the position of the kth sampling region in the convolution kernel, JPEG2025181612000011.jpg711 represents the offset of the sampling region, allowing the shape of the convolution kernel to be adaptively adjusted.

[0012] In the above small 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, and a two-layer spatial and channel reconstruction convolution SCConv is introduced in the middle bottleneck layer of the C2f_SCConv module, which is composed of a spatial reconstruction module SRU and a channel reconstruction module CRU.

[0013] In the above small vehicle detection deep learning model based on multi-scale module feature fusion, the formula for GroundNorm in the spatial reconstruction module SRU is as follows:

[0014]

number

[0015] where: X is the input vector, μ is the mean value of X, σ is the standard deviation of X, ε is 1×10 -10 and β is a trainable parameter, γ is a trainable parameter, JPEG2025181612000013.jpg514 is a trainable affine transform, By performing a standard normalization operation on JPEG2025181612000014.jpg54, the information weight The result is JPEG2025181612000015.jpg615, A threshold-based selection operation is performed on JPEG2025181612000016.jpg45, and the weights with the highest information content are JPEG2025181612000017.jpg78 and weights with less information We take JPEG2025181612000018.jpg610 and pass the input through these weights JPEG2025181612000019.jpg76 and Take JPEG2025181612000020.jpg75, divide it into two parts based on the channels, multiply each part alternately, and then concatenate them based on the channels.

[0016] In the above-mentioned small vehicle detection deep learning model based on multi-scale module feature fusion, the channel reconstruction module CRU performs the following operations on the input data: After performing the segmentation based on JPEG2025181612000021.jpg75, the upper layer convolution operation and the lower layer convolution operation are performed respectively, where the upper layer convolution operation is JPEG2025181612000022.jpg612 and It contains the convolution kernel of JPEG2025181612000023.jpg512, and the underlying convolution operation is It includes a convolution kernel of JPEG2025181612000024.jpg511 and a simple skip-connection, followed by channel concatenation based on average pooling and softmax functions.

[0017] In the above-mentioned small vehicle detection deep learning model based on multi-scale module feature fusion, the SPPF_LSKA module passes through the first convolution layer Conv, then passes through three max pooling layers, and concatenates the outputs obtained by max pooling in each layer. Then it passes through the large separable convolution kernel attention module LSKA and the convolution layer Conv to obtain the output.

[0018] In the above-mentioned small vehicle detection deep learning model based on multi-scale module feature fusion, the large separable convolutional kernel attention module (LSKA) is constructed by multiple convolutional layers (Conv), each of which has a different convolutional kernel. The horizontal convolutional kernels are arranged in order. JPEG2025181612000025.jpg624, vertical convolution kernel JPEG2025181612000026.jpg524, horizontal convolution kernel JPEG2025181612000027.jpg617, vertical convolution kernel JPEG2025181612000028.jpg616, JPEG2025181612000029.jpg512 is the convolution kernel in units of 12, where k represents the original convolution kernel size and d denotes the dilation factor of the convolution.

[0019] In the above-mentioned small vehicle detection deep learning model based on multi-scale module feature fusion, the multi-scale convolutional kernel detection module MSK_Detect performs a channel separation operation through a separation operation, and the output obtained by the separation operation passes through a convolutional layer Conv with multiple convolutional kernels of different sizes, and then performs decoupling through a convolutional layer Conv with a 1x1 convolutional kernel, and then uses two 1x1 convolutional layers Conv respectively to obtain the object bounding box and object classification output results.

[0020] In the above small vehicle detection deep learning model based on multi-scale modular feature fusion, this small vehicle object detection deep learning model is improved based on the YOLOv8 model as the object detection model. [Effects of the Invention]

[0021] Compared with conventional techniques, the advantages of this invention are as follows: The introduction of the C2f_DCNv3 module into the backbone network achieves rich feature extraction capabilities, providing a powerful and efficient backbone network foundation for the entire model, ensuring high performance and robustness of the overall model when processing image feature extraction tasks for small vehicles. The backbone network incorporates the SPPF_LSKA module into its deep structure. The SPPF_LSKA module fuses deep receptive fields of different scales while capturing more long-range dependencies in the feature maps of small vehicles based on the large separable kernel attention module LSKA, achieving better spatial and channel adaptability. The combination of module construction offers advantages such as high efficiency, ease of integration, optimized parameter configuration, and reduced parameter amount, thereby reducing computational and storage costs. The C2f_SCConv module is introduced into the backbone layer. The C2f_SCConv module offers advantages such as high efficiency, ease of integration, optimized parameter configuration, and reduced parameter amount, reducing computational and storage costs while providing better performance than traditional convolutional operations. A multi-scale convolution kernel detection module MSK_Detect is introduced into the head network to improve the diversity of image feature representation of small vehicles and strengthen the generalization ability of the network when faced with complex features; at the same time, the design of this multi-scale convolution kernel detection module MSK_Detect can reduce the number of parameters. [Brief explanation of the drawings]

[0022] [Figure 1] FIG. 1 is an overall architectural diagram of the model of the present invention. [Figure 2] FIG. 1 is a structural schematic diagram of the C2f_DCNv3 module and the deformable convolutional DCNv3 of the present invention. [Figure 3] FIG. 1 is a structural schematic diagram of the C2f_SCConv module and the spatial channel reconstruction convolution SCConv of the present invention; [Figure 4]1 is a structural schematic diagram of the SPPF_LSKA module and the large separable convolution kernel attention module LSKA of the present invention; [Figure 5] FIG. 2 is a structural schematic diagram of the multi-scale convolution kernel detection module MSK_Detect of the present invention; [Figure 6] 1 is a schematic diagram of a small vehicle detection comparison of the present invention; DETAILED DESCRIPTION OF THE INVENTION

[0023] The present invention will now be described in more detail with reference to the drawings and specific embodiments.

[0024] As shown in Figure 1, this is a deep learning model for small vehicle detection based on multi-scale modular feature fusion. This model can improve the safety of autonomous driving by helping 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 improve the efficiency of security monitoring by helping monitoring systems quickly and accurately identify small objects such as suspicious persons and vehicles, thereby better protecting public safety. This small vehicle detection deep learning model includes a backbone network, a neck layer, and a head network. The backbone network includes a C2f_DCNv3 module (deformable convolution and double convolution cross-stage module) based on the combination of a deformable convolution DCNv3 and a double convolution cross-stage module C2f (cross-stage feature fusion), and a SPPF_LSKA module (spatial pyramid pooling layer and large separable convolution kernel attention module) based on the combination of a spatial pyramid pooling layer SPPF (Spatial Pyramid Pooling-Fast) and a large separable convolution kernel attention module LSKA (Large Separable Kernel Attention) to extract features from small vehicle images. The neck layer includes a C2f_SCConv module (spatial channel reconstruction convolution and double convolution cross-stage module) based on the combination of a 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 includes a multi-scale convolution kernel detection module MSK_Detect (Multi Scale Kernel_Detect (a multi-scale convolution kernel detection head) is introduced to predict the extracted image features of small vehicles. The above small vehicle detection deep learning model can efficiently handle the task of detecting small vehicle objects in complex images.

[0025] Figure 2 shows the overall structure of the C2f_DCNv3 module (deformable convolution and double convolution cross-stage module), which is based on the combination of the deformable convolution DCNv3 and the double convolution cross-stage module C2f, respectively. The left side of Figure 2 shows the propagation structure of the C2f_DCNv3 module. First, the input and output of the C2f_DCNv3 module pass through the convolution layer Conv. A two-layer deformable convolution DCNv3 is introduced in the middle bottleneck layer of the C2f_DCNv3 module, where various types of multi-layer modules can be introduced. The right side of Figure 2 explains the structure of the deformable convolution DCNv3. A sampling grid (sampling_grid) is obtained through the propagation operation of tensor operations based on the offset weight offset (offset) and scalar weight mask (mask). This is used as the sampling operation domain by the deformable convolution DCNv3, and finally the output of the deformable convolution DCNv3 is obtained.

[0026] In detail, the specific calculation process of deformable convolution DCNv3 is as follows:

[0027]

number

[0028] where: JPEG2025181612000031.jpg66 is the feature map JPEG2025181612000032.jpg56 indicates the output result calculated by DCNv3 calculation. JPEG2025181612000033.jpg55 represents the number of aggregation groups, and DCNv3 uses multiple spatial locations to learn multiple spatial aggregation modes. JPEG2025181612000034.jpg64 represents the total number of sampling areas, JPEG2025181612000035.jpg66 represents the weight of the convolution kernel in the sampling operation, JPEG2025181612000036.jpg623 represents the modulation scalar, which is the normalized coefficient for adjusting the contribution weight of each sampling region, JPEG2025181612000037.jpg77 represents the input feature map, JPEG2025181612000038.jpg76 represents the position of the kth sampling region in the convolution kernel, JPEG2025181612000039.jpg711 represents the offset of the sampling region, allowing the shape of the convolution kernel to be adaptively adjusted.

[0029] The constructed C2f_DCNv3 module, as an important part of the backbone network of the model of the present invention, possesses powerful feature extraction capabilities. By incorporating long-distance dependencies and adaptive spatial aggregation mechanisms, the module efficiently handles large-scale image data, thereby learning more powerful and robust feature representations. The C2f_DCNv3 module has higher parameter utilization efficiency, and by incorporating weight sharing and multi-group mechanisms, it can achieve rich feature extraction capabilities while ensuring parameter and computational efficiency. Furthermore, the C2f_DCNv3 module exhibits excellent stability during the training process, effectively mitigating instability factors that may arise during training. This improved stability not only optimizes the training process, but also improves the model's generalization ability and reliability. Thus, the addition of the C2f_DCNv3 module provides a powerful and efficient backbone network foundation for the model of the present invention, ensuring the overall high performance and robustness of the model when processing complex vision tasks.

[0030] Figure 3 shows the structure of the C2f_SCConv module, which is based on the combination of the spatial channel reconstruction convolutional SCConv and the double convolutional cross-stage module C2f, and the overall structure of the spatial channel reconstruction convolutional SCConv module. The left part of Figure 3 shows the propagation structure of the C2f_SCConv module, which is similar to the C2f framework of the C2f_DCNv3 module. The input and output of the C2f_SCConv module pass through the convolutional layer Conv, and a two-layer spatial and channel reconstruction convolutional SCConv is introduced in the middle bottleneck layer of the C2f_SCConv module. The spatial and channel reconstruction convolutional SCConv consists of a spatial reconstruction module SRU and a channel reconstruction module CRU.

[0031] Furthermore, the central part of FIG. 3 shows the structure of the spatial reconstruction module SRU, and the formula of GroundNorm in the spatial reconstruction module SRU is as follows:

[0032]

number

[0033] where: X is the input vector, μ is the mean value of X, σ is the standard deviation of X, ε is 1×10 -10 and β is a trainable parameter, γ is a trainable parameter, JPEG2025181612000041.jpg514 is a trainable affine transform, By performing a standard normalization operation on JPEG2025181612000042.jpg54, the information weight The result is JPEG2025181612000043.jpg614, A threshold-based selection operation (Gate operation) is performed on JPEG2025181612000044.jpg45, and the weight with the highest information content is selected. JPEG2025181612000045.jpg78 and weights with less information We take JPEG2025181612000046.jpg610 and pass the input through these weights JPEG2025181612000047.jpg76 and JPEG2025181612000048.jpg75 is taken, divided into two parts based on the channel, and then multiplied alternately, and then concatenated based on the channel. JPEG2025181612000049.jpg45 is sorted for each value, and those with values ​​of 0.5 or more are selected. JPEG2025181612000050.jpg66 value JPEG2025181612000051.jpg79, and less than 0.5 The value of JPEG2025181612000052.jpg66 is JPEG2025181612000053.jpg611. A standard normalization operation is, for example, JPEG2025181612000054.jpg1325.

[0034] In addition, the right part of Figure 3 shows the structure of the channel reconfiguration module CRU, which performs the following operations on the input data: After performing the segmentation based on JPEG2025181612000055.jpg75, the upper layer convolution operation and the lower layer convolution operation are performed respectively, where the upper layer convolution operation is JPEG2025181612000056.jpg612 and It contains the convolution kernel of JPEG2025181612000057.jpg512, and the underlying convolution operation is It includes a convolution kernel of JPEG2025181612000058.jpg511 and a simple skip connection, and then performs channel concatenation based on the average pooling and softmax functions.

[0035] The spatial channel reconstruction convolutional SCConv enhances the information flow between features through separation and reconstruction and cross reconstruction methods, thereby improving feature extraction efficiency and feature representation ability and reducing spatial redundancy. The spatial channel reconstruction convolutional SCConv can more accurately capture and represent important information in images, especially when processing complex and subtle features. In this invention, the C2f_SCConv module is introduced into the bottleneck layer. The C2f_SCConv module has the advantages of high efficiency, easy integration, optimized parameter configuration, and reduced parameter amount, reducing computational and storage costs while providing better performance than traditional convolution operations.

[0036] As shown in Figure 4, the overall structure of the SPPF_LSKA module, which is constructed by combining the large separable kernel attention module LSKA and the spatial pyramid pooling layer SPPF, and the internal structure of the large separable kernel attention module LSKA are shown. The left part of Figure 4 shows the propagation structure of the SPPF_LSKA module, in which the SPPF_LSKA module passes through the first convolution layer Conv, then goes through three max pooling layers, concatenates the outputs obtained from each max pooling layer, and then passes through the large separable kernel attention module LSKA and the convolution layer Conv to obtain the output.

[0037] As shown in Figure 4, the right part of Figure 4 shows the internal structure of the large separable kernel attention module LSKA. The large separable convolution kernel attention module LSKA is constructed by multiple convolution layers Conv, each of which has a different convolution kernel. The horizontal convolution kernels JPEG2025181612000059.jpg624, vertical convolution kernel JPEG2025181612000060.jpg524, horizontal convolution kernel JPEG2025181612000061.jpg617, vertical convolution kernel JPEG2025181612000062.jpg615, JPEG2025181612000063.jpg512 is the convolution kernel in units of 12, where k represents the original convolution kernel size and d denotes the dilation factor of the convolution.

[0038] LSKA aims to solve the problem of deep convolutional layers' computational and memory usage increasing linearly with the size of the convolution kernel. LSKA decomposes 2D convolution kernels into horizontal and vertical 1D convolution kernels, enabling deep convolutional layers with large convolution kernels to be directly used without additional modules. LSKA decomposes the 2D weight kernels of deep convolution and deep dilated convolution into two serially concatenated 1D separable weight kernels, reducing computational complexity while maintaining performance comparable to conventional large convolutional kernel modules.

[0039] The advantages of the large separable convolution kernel attention module LSKA are as follows: 1. Regarding long-range dependencies: LSKA effectively captures long-range dependencies within an image, which is an important component of the self-attention mechanism implemented in the transformer. LSKA provides an alternative method for capturing global information through the design of large convolution kernels. 2. Regarding spatial and channel adaptability: LSKA inherits the spatial attention and channel attention strategies of traditional large convolution kernels. These two strategies adaptively readjust feature weights based on contextual dependencies. Compared with self-attention and LKA modules, LSKA has better spatial and channel adaptability while reducing the number of parameters and computational cost. 3. Regarding computational complexity and memory usage: Compared to standard traditional large convolution kernel modules, LSKA can significantly reduce computational complexity and memory usage.

[0040] In this invention, the SPPF_LSKA module is introduced into the deep structure of the backbone network to fuse deep receptive fields of different scales, and at the same time, based on the LSKA module, it can capture more long-distance dependencies within features and achieve better spatial and channel adaptability.

[0041] The structure of the multi-scale convolution kernel detection head MSK_Detect is shown in Figure 5. The multi-scale convolution kernel detection module MSK_Detect uses a split operation to separate channels, and the output obtained by the split operation passes through a convolution layer Conv with multiple convolution kernels of different sizes. In this invention, two types of convolution kernels, 3x3 and 5x5, are selected to form the multi-scale convolution kernel, and then pass through a convolution layer Conv with a 1x1 convolution kernel for separation. After that, two 1x1 convolution layers Conv are used to obtain the object bounding box and object classification output results.

[0042] The MSK module's multi-scale convolution kernels enable it to capture features at different levels, improving the diversity of feature representations and strengthening the network's generalization ability for complex features, allowing the model to demonstrate 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 and optimizes the use and fusion of convolution kernels, thereby optimizing the parameter calculation efficiency of the MSK module and reducing computational costs and memory usage. In this invention, the MSK module is introduced into the head layer to construct the multi-scale convolution kernel detection module MSK_Detect. By applying multi-scale convolution kernels, the MSK_Detect module effectively improves its ability to capture and process features of different sizes related to small vehicle objects. This multi-scale approach allows the module to understand the content of small vehicles at different levels, thereby improving the accuracy and robustness of small vehicle detection.

[0043] Preferably, this small object detection deep learning model is improved based on the YOLOv8 model as an object detection model, and the effectiveness of this method is verified by the following comparative experiments.

[0044] A deep learning experiment was conducted based on VisDrone, a large-scale real-world scene image dataset containing small vehicles. The samples in this dataset are diverse and abundant, encompassing multiple vehicle categories. A small object detection model based on the YOLOv8 algorithm and the proposed multi-scale modular feature fusion was trained and tested on the test set. Experimental results evaluated the model's performance based on the mean average precision (mAP) evaluation metric on the test dataset. The YOLOv8 model achieved an mAP50 of 49.7% and an overall mAP of 30.3%, while the proposed model achieved an mAP of 52.5% and an overall mAP of 32.1%. These results demonstrate that the proposed model outperforms the traditional YOLOv8 method in detecting small vehicle objects in large-sized images of real scenes.

[0045] The following table shows the comparison results of the performance metrics of the YOLOv8 model and the small vehicle detection deep learning model based on multi-scale modular feature fusion proposed in this application on the small vehicle object detection dataset, demonstrating the performance improvement of the method proposed in this application.

[0046] [Table 1]

[0047] Figure 6 shows a comparative example of small vehicle object detection results in real scenes, where Figure 6(a) shows the detection results using the conventional YOLOv8 model, and Figure 6(b) shows the detection results using the newly proposed model in this application. The model in this application can detect small vehicle objects better than the YOLOv8 model.

[0048] From the above, the principle of this embodiment is as follows: based on the overall structure, improvements are made in many aspects to build a deep network feature detection head for the size of small object vehicles, make the model more adapted to learning the feature size of small object vehicles, introduce a deformable convolutional DCNv3 into the backbone network to improve the processing ability and stability of input features, introduce a large separable convolutional kernel attention module LSKA into the deep stage of the backbone network to better capture the overall information of the features, and introduce a spatial channel reconstruction convolutional SCConv into the neck layer to effectively reduce the redundancy of the convolutional network, while at the same time improving the performance and efficiency of the network, and achieving better identification and location effects for small vehicle objects.

[0049] The specific examples described herein are merely illustrative of the spirit of the present invention, and those skilled in the art may make various modifications, additions, or substitutions in similar ways to the specific examples described, without departing from the spirit of the present invention or beyond the scope defined by the appended claims.

[0050] In this specification, terms such as backbone network, neck layer, and head network are frequently used, but this does not exclude the possibility that other terms may be used. These terms are used only to more conveniently explain and interpret the essence of the present invention, and interpreting them as any additional restrictions would deviate from the spirit of the present invention.

Claims

1. A small vehicle detection deep learning model based on multi-scale modular feature fusion, including a backbone network, a neck layer, and a head network, The backbone network introduces a C2f_DCNv3 module based on a combination of a deformable convolutional DCNv3 and a double convolutional cross-stage module C2f, and an SPPF_LSKA module based on a combination of a spatial pyramid pooling layer SPPF and a large separable convolutional kernel attention module LSKA to extract features from images of small vehicles; introducing a C2f_SCConv module based on a combination of a spatial channel reconstruction convolution SCConv and a double convolution cross-stage module C2f into the neck layer to process the features extracted by the backbone network; A small vehicle detection deep learning model based on the fusion of multi-scale module features, characterized by introducing a multi-scale convolution kernel detection module MSK_Detect into the head network and making predictions on the extracted image features of small vehicles.

2. The small vehicle detection deep learning model based on multi-scale modular feature fusion according to claim 1, characterized in that the input and output of the C2f_DCNv3 module pass through a convolutional layer Conv, and a two-layer deformable convolutional DCNv3 is introduced as a middle bottleneck layer of the C2f_DCNv3 module.

3. The specific calculation process of the deformable convolution DCNv3 is as follows: [Equation 1] where: 【number】 is the feature map 【number】 The value at the position indicates the output result calculated by the DCNv3 operation, 【number】 represents the number of aggregation groups, and represents learning multiple spatial aggregation modes using multiple spatial locations in DCNv3 operation; 【number】 represents the total number of sampling areas, 【number】 represents the weight of the convolution kernel in the sampling operation, 【number】 represents the modulation scalar, which is a normalized coefficient for adjusting the contribution weight of each sampling region, 【number】 represents the input feature map, 【number】 represents the position of the kth sampling region in the convolution kernel, 【number】 The deep learning model for small vehicle detection based on multi-scale modular feature fusion according to claim 2, wherein represents the offset amount of the sampling region, and the shape of the convolution kernel can be adaptively adjusted.

4. 2. The small vehicle detection deep learning model based on multi-scale modular feature fusion according to claim 1, wherein the input and output of the C2f_SCConv module pass through a convolution layer Conv, and a two-layer spatial and channel reconstruction convolution SCConv is introduced in the middle bottleneck layer of the C2f_SCConv module, and the spatial and channel reconstruction convolution SCConv is composed of a spatial reconstruction module SRU and a channel reconstruction module CRU.

5. The formula of GroundNorm in the spatial reconstruction module SRU is as follows: [Equation 2] where: X is the input vector, μ is the mean value of X, σ is the standard deviation of X, ε is 1 × 10 －１０ and β is a trainable parameter, γ is a trainable parameter, 【number】 is a trainable affine transformation, 【number】 By performing a standard normalization operation on 【number】 Get 【number】 A threshold-based selection operation is performed on the weights with high information content, 【number】 and weights with less information 【number】 and the input is passed through these weights 【number】 and 【number】 The deep learning model for small vehicle detection based on multi-scale modular feature fusion according to claim 4, characterized in that:

6. The channel reconfiguration module CRU After performing the segmentation based on, the upper layer convolution operation and the lower layer convolution operation are performed respectively, and the upper layer convolution operation is and The convolution kernel is included, and the underlying convolution operation is The small vehicle detection deep learning model based on multi-scale modular feature fusion according to claim 4, characterized in that it includes a convolution kernel and a simple skip connection, and then performs channel concatenation based on an average pooling operation and a softmax function operation.

7. The small vehicle detection deep learning model based on multi-scale module feature fusion according to claim 1, characterized in that the SPPF_LSKA module passes through the first convolutional layer Conv, then passes through three max pooling layers, and concatenates the outputs obtained by the max pooling operations of each layer, and then passes through the large separable convolutional kernel attention module LSKA and the convolutional layer Conv to obtain the output.

8. The large separable convolution kernel attention module LSKA is constructed by a plurality of convolution layers Conv, each of which has a different convolution kernel. , the vertical convolution kernel , horizontal convolution kernel , the vertical convolution kernel 、 The small vehicle detection deep learning model based on multi-scale modular feature fusion according to claim 7, characterized in that the convolution kernel is in units of k, k represents the original convolution kernel size, and d represents the convolution expansion coefficient.

9. 2. The small vehicle detection deep learning model based on multi-scale module feature fusion according to claim 1, wherein the multi-scale convolution kernel detection module MSK_Detect performs a channel separation operation through a separation operation, and the output obtained by the separation operation passes through a convolution layer Conv of multiple convolution kernels of different sizes, and then performs decoupling through a convolution layer Conv of a 1x1 convolution kernel, and then obtains an object bounding box and object classification output results using two 1x1 convolution layers Conv respectively.

10. The small vehicle detection deep learning model based on multi-scale modular feature fusion as claimed in claim 1, characterized in that the small vehicle detection deep learning model object is improved based on the YOLOv8 model as the object detection model.

Images