Low-altitude aerial photography vehicle small target detection method based on deep learning

Abstract

The invention relates to the technical field of target detection, in particular to a low-altitude aerial vehicle small target detection method based on deep learning, and the method comprises the steps: obtaining a low-altitude aerial image, inputting the low-altitude aerial image into a detection model for processing, and outputting a target detection result; the processing steps are realized through the following modules in sequence: a multi-branch feature fusion module used for extracting and fusing context information and fine-grained information under different receptive fields to obtain an initial fusion feature map; a multi-branch convolution attention module connected to the multi-branch feature fusion module and used for performing attention weighting of channel and spatial dimensions on the initial fusion feature map to obtain an enhanced intermediate feature map; and the reverse enhancement module is connected to the multi-branch convolution attention module and is used for carrying out weighted regulation and control according to the deep semantic features contained in the intermediate feature map and outputting a target detection result. The accuracy of the target detection result determined through the detection model is higher.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, specifically to a method for detecting small targets of vehicles in low-altitude aerial photography based on deep learning. Background Technology

[0002] In recent years, with the popularization of drone technology, real-time vehicle detection using its low-altitude perspective and high maneuverability has become a key technology in fields such as intelligent transportation, public safety, and disaster response. Low-altitude aerial photography can acquire high-resolution, detailed images, providing a data foundation for accurate identification of target vehicles, and also placing higher demands on the accuracy and efficiency of detection algorithms.

[0003] Currently, deep learning models based on Convolutional Neural Networks (CNNs) are the mainstream approach. However, when these general detection models are directly applied to low-altitude aerial photography scenarios, their performance often degrades due to the unique characteristics of imaging, mainly facing the following three challenges: variable target scale, complex background information, and a large number of small targets.

[0004] First, the diversity of target scales leads to significant size differences among similar targets in images, affecting the consistency of feature extraction and weakening the model's recognition ability. Existing methods often improve multi-scale perception capabilities through Feature Pyramid Networks (FPNs) or multi-scale contextual information. However, conventional FPN structures have long information flow paths, and sparse features of small targets are easily diluted or lost during transmission. At the same time, simple feature fusion operations are difficult to effectively reconcile the semantic differences between shallow and deep features.

[0005] Secondly, low-altitude aerial images often contain a large amount of complex background, whose texture, color, and other features are highly similar to the target, which can easily interfere with the model. Although existing research has introduced attention mechanisms or contextual modeling to improve the focus on the target region, the background pixels account for a much larger proportion in aerial images, causing the global information to be dominated by the background, making it difficult to accurately focus on the target, and introducing additional computational overhead.

[0006] Furthermore, aerial images contain numerous small targets with low pixel counts, whose features are weak, easily obscured by the background, and lost during network transmission. Some studies have attempted to improve expressive power using super-resolution reconstruction or small target feature enhancement modules. However, super-resolution reconstruction, as a preprocessing technique, may introduce unrealistic texture artifacts, interfering with the detection task, and significantly increasing the computational burden, making it difficult to meet the real-time requirements of UAV platforms.

[0007] In summary, existing aerial target detection methods either address only a single problem or improve performance by stacking complex modules, often neglecting the inherent connections between multiple issues and struggling to achieve a good balance between detection accuracy and computational efficiency. Therefore, there is an urgent need for a new method that can efficiently and collaboratively solve the challenges of detecting scale variations, background interference, and small targets. Summary of the Invention

[0008] To address the aforementioned technical problem of low performance in existing target detection models applied to low-altitude aerial photography, the present invention aims to provide a deep learning-based method for detecting small vehicle targets in low-altitude aerial photography. The specific technical solution adopted is as follows: One embodiment of the present invention provides a method for detecting small targets of vehicles in low-altitude aerial photography based on deep learning, the method comprising the following steps: A low-altitude aerial image containing the vehicle is acquired, and the image is input into a detection model for processing to output the target detection result. The processing steps are implemented sequentially through the following modules: The multi-branch feature fusion module is used to extract and fuse contextual information and fine-grained information from different receptive fields based on the low-altitude aerial image to obtain an initial fused feature map. A multi-branch convolutional attention module, connected to the multi-branch feature fusion module, is used to perform channel and spatial dimension attention weighting on the initial fused feature map to obtain an enhanced intermediate feature map; The top-down reverse enhancement module is connected to the multi-branch convolutional attention module. It is used to generate guiding signals based on the deep semantic features contained in the intermediate feature map, perform weighted regulation, and output the target detection result.

[0009] Further, obtaining the initial fused feature map includes: The multi-branch feature fusion module includes multiple convolutional branches set in parallel. Each convolutional branch has a receptive field of different size and is configured to perform directional modeling to extract local contextual features, global contextual features, and fine-grained directional information representing target edges and contours from the low-altitude aerial image, respectively. The multi-branch feature fusion module is further configured to fuse the local context features, global context features, and fine-grained directional information extracted from each of the convolutional branches to generate the initial fused feature map.

[0010] Furthermore, the convolutional branch includes: The upper branch is used to enhance the model's ability to distinguish between the core target area and the local background through center-surround contrast modeling; The middle branch is used to enhance the model's sensitivity to the directional features of the target contour by utilizing asymmetric strip convolutions through directional modeling. The next branch is used to collaboratively capture the local details and global semantic information of the target through multi-scale context modeling, so as to enhance the understanding of the target's contextual dependencies.

[0011] Furthermore, the multi-branch feature fusion module also includes: The feature segmentation unit is used to decouple the input features according to their functions, retaining local gradient information and high-level semantic information respectively; Cascaded units are used to perform multi-level semantic enhancement and feature extraction on the semantic features extracted by the feature segmentation unit. The feature fusion unit is used to integrate features from different paths to obtain the initial fused feature map, which contains both shallow details and deep semantics.

[0012] Furthermore, the multi-branch convolutional attention module includes: A multi-branch feature extraction unit is used to process the input initial fused feature map to generate a basic feature map containing multi-scale contextual information. The channel attention branch is used to generate channel attention weights based on the global information of the base feature map; A spatial attention branch is used to generate a spatial attention map based on the spatial distribution of the base feature map. And a feature refinement unit, used to jointly weight the basic feature map using the channel attention weights and the spatial attention map to generate the enhanced intermediate feature map.

[0013] Further, the output of the target detection result includes: A reverse attention mask is generated based on the deep semantic features, serving as the guiding signal; The intermediate feature map is weighted using the guiding signal to obtain a weighted feature map. The weighted feature map is subjected to channel compression processing, and the target detection result is determined based on the result of the channel compression processing.

[0014] Furthermore, the overall architecture of the detection model is YOLOv8.

[0015] Furthermore, the detection model is deployed on an embedded platform of an unmanned aerial vehicle (UAV).

[0016] The present invention has the following beneficial effects: This invention provides a deep learning-based method for detecting small targets of vehicles in low-altitude aerial photography. Using YOLOv8 as the baseline model, this method proposes a model called MTNet for detecting small targets of vehicles in low-altitude aerial photography. Its core lies in the integration of three efficient plug-and-play modules, achieving powerful multi-branch contextual feature fusion and reverse enhancement functions. First, a multi-branch contextual feature fusion module (C2f-MFE) is designed, employing a three-branch parallel convolutional structure to fuse multi-scale receptive field information, enriching the local and contextual feature representation of the target and improving the model's robustness to scale changes. Second, a multi-branch convolutional attention mechanism (MCAM) is designed, using a combination of channel and spatial attention to guide the model to adaptively focus on the target region, effectively suppressing the interference of complex background noise and enhancing the response of weak target features. Finally, a top-down reverse enhancement module (TDRE) is constructed, utilizing deep semantic information to generate an attention mask, which is applied to shallow features to enhance shallow small target features, compensating for information loss during network transmission of small targets, thereby effectively mitigating the problem of missed detection. Attached Figure Description

[0017] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the steps of a method for detecting small targets in low-altitude aerial photography vehicles based on deep learning, as an embodiment of the present invention. Figure 2 This is a schematic diagram of the detection model in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the multi-branch feature extraction module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the multi-branch feature fusion module in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the multi-branch convolutional attention module in an embodiment of the present invention; Figure 6 This is a schematic diagram of the top-down reverse enhancement module in an embodiment of the present invention; Figure 7 This is a thermal effect diagram of TDRE in an embodiment of the present invention; Figure 8 This is a visualization comparison of the target detection results of MTNet in the DroneVehicle dataset in an embodiment of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the technical solution proposed according to the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] To address the limitations of existing technologies in detecting multi-scale variations, complex backgrounds, and small targets, and to overcome their inability to coordinate multiple challenges and maintain efficiency, this invention proposes a solution: constructing a unified detection framework that can efficiently and collaboratively resolve these shortcomings. Based on in-depth research into the problem, three innovative modules that can mutually promote and complement each other are specifically designed and integrated into the advanced YOLOv8 architecture, forming the detection model proposed in this invention, namely the MTNet model.

[0022] To achieve accurate and efficient detection of small vehicle targets in low-altitude aerial images, one embodiment of the present invention provides a deep learning-based method for detecting small vehicle targets in low-altitude aerial images, the method comprising the following steps: Acquire low-altitude aerial images containing vehicles, input these images into a detection model for processing, and output target detection results.

[0023] As an exemplary implementation, such as Figure 1 As shown, the steps to output the target detection results may include: S101: Acquire low-altitude aerial images containing vehicles and preprocess the images.

[0024] In this embodiment, raw low-altitude aerial images containing vehicles are acquired using an image sensor (such as a visible light camera) mounted on a drone platform. The drone cruises within a specific mission area (such as urban roads or transportation hubs) to ensure that the acquired images have a low-altitude perspective, high resolution, and rich scene details.

[0025] After acquiring the original image, a series of standardized preprocessing operations are required to adapt it to the input requirements of subsequent deep learning models and improve processing efficiency and stability. The preprocessing steps include: Size normalization: The original image is scaled or cropped to a fixed size specified by the model (e.g., 640×640 pixels) to ensure consistency in batch processing.

[0026] Color space normalization: Converting and normalizing an image from its original RGB color space, for example, normalizing pixel values ​​to a color space of 10 ... The interval can be standardized (subtract the mean, divide by the standard deviation) to accelerate model training convergence and improve generalization ability.

[0027] Data augmentation: To improve the robustness of the model, online data augmentation techniques such as random horizontal flipping, random brightness or contrast adjustment, and Mosaic stitching can be used during training to simulate different imaging conditions and enrich the diversity of training samples.

[0028] It is worth noting that in this embodiment, the pre-processed image will be used as the input data to be detected.

[0029] S102, input the preprocessed image into the pre-trained detection model for processing, and output the target detection result.

[0030] In this embodiment, the overall architecture of the detection model (MTNet model) is based on YOLOv8 and integrates three efficient and pluggable modules to build a new framework more suitable for low-altitude aerial small target detection. The three modules are a multi-branch feature fusion module 10 (C2f-MFE), a multi-branch convolutional attention module 11 (MCAM), and a top-down reverse enhancement module 12 (TDRE). The structural diagram of the detection model is shown below. Figure 2 As shown, the three core modules of the MTNet model are specifically deployed in the backbone and neck of YOLOv8.

[0031] The detection model's processing steps are implemented sequentially through the following modules to extract and refine deep features from the input image: The multi-branch feature fusion module 10 is used to extract and fuse contextual information and fine-grained information under different receptive fields based on low-altitude aerial images to obtain an initial fused feature map.

[0032] Here, the multi-branch feature fusion module aims to improve the expressive power for multi-scale targets and the feature perception capability for small targets by designing multi-scale convolutional branches to construct a hierarchical receptive field structure. The structural diagram of the multi-branch feature fusion module is shown below. Figure 4 As shown, in Figure 4In the MFE (Multi-Branch Feature Fusion) module, the specific structure consists of three branches (upper, middle, and lower) and a feature fusion layer. The upper, middle, and lower branches construct three branch paths with complementary receptive fields, focusing on center contrast, direction awareness, and context modeling, respectively, enriching feature representations from multiple perspectives. It's worth noting that the feature representations extracted from the three branches—i.e., multi-scale semantic information—can be effectively integrated in the fusion layer, significantly improving the consistency and expressive power of the overall features and providing a more discriminative feature foundation for small object detection.

[0033] To alleviate the performance degradation caused by drastic scale changes and occlusion in small target detection, numerous studies have explored multi-scale contextual relationships to improve target recognition. Existing research indicates that the rational utilization of multi-scale contextual information can significantly improve target recognition performance under complex backgrounds and occlusion conditions, laying a theoretical foundation for subsequent context-based small target detection methods. In UAV scenarios, Zhao et al. proposed the IUAV-YOLO detector, which achieves joint modeling of multi-scale local and global contextual information by introducing a multi-branch backbone feature extraction module and a spatial context-aware module (SCAM). The multi-branch backbone module employs a parallel convolutional structure with different kernel sizes and dilation rates, while SCAM combines global max pooling and global average pooling, and models cross-channel and cross-spatial dependencies through matrix interaction, thereby enhancing the response capability of small targets in complex backgrounds. Although this method achieves a significant improvement in small target detection accuracy, the superposition of multi-branch convolution, global pooling, and attention mechanisms inevitably increases model complexity and computational overhead, limiting its application on real-time UAV platforms. From another perspective, Fang et al. proposed a global-local dilated residual network to highlight small targets in UAV images by predicting foreground-background residual maps. This method alternately stacks globally dilated residual blocks with large receptive fields and locally dilated residual blocks with smaller dilation rates, thereby simultaneously modeling long-range contextual relationships and fine-grained edge information, effectively enhancing the response to small targets under low-contrast conditions. However, this method relies on a multi-layered dilated residual structure, resulting in high computational overhead and difficulty in seamless integration into standard single-stage detection frameworks, hindering lightweight deployment. The aforementioned research demonstrates that multi-branch convolutional structures and carefully designed dilated residual modules have significant advantages in multi-scale contextual modeling, but also reveal shortcomings in real-time performance and engineering adaptability. Based on this, this embodiment proposes a multi-branch contextual feature fusion module (C2f-MFE). This module, through a parallel multi-branch structure, collaboratively models contextual and fine-grained information under different receptive fields without significantly increasing computational burden, enhancing the complementarity between features and thus more efficiently and robustly addressing the challenges posed by drastic scale changes.

[0034] In this embodiment, the multi-branch feature fusion module includes multiple convolutional branches set in parallel. Each convolutional branch has a receptive field of different size and is configured to perform directional modeling to extract local contextual features, global contextual features, and fine-grained directional information representing the edges and contours of the target from the low-altitude aerial image. The multi-branch feature fusion module is also configured to fuse the local contextual features, global contextual features, and fine-grained directional information extracted by each convolutional branch to generate an initial fused feature map.

[0035] Furthermore, the C2f-MFE module replaces all C2f modules in the Backbone. Its input is the feature map output from the previous convolutional downsampling module (CBS), and its output is a more expressive feature map enhanced by C2f-MFE. To address the issue of frequent target scale changes, C2f-MFE introduces a multi-branch feature extraction module (MFE) while maintaining the C2f topology. A schematic diagram of the multi-branch feature extraction module is shown below. Figure 3 As shown, the C2f-MFE module consists of a feature segmentation unit, a cascade unit, and a feature fusion unit. The feature segmentation unit is responsible for dividing the input features into multiple groups of sub-features, preserving multi-layer semantic flow to maintain information integrity; the cascade unit introduces a three-branch MFE structure between feature channels to perform center-surround contrast modeling, directionality modeling, and multi-scale modeling, respectively, to achieve multi-dimensional enhancement of space, semantics, and receptive field; the feature fusion unit integrates and fuses the multi-branch outputs through channel concatenation and convolution to generate a feature map with stronger representational capabilities, i.e., the initial fused feature map.

[0036] For the upper branch 101 of the multi-branch feature extraction module (MFE), the aim is to enhance the model's ability to distinguish between the target core region and the local background through center-surround contrast modeling. The implementation process includes: first, inputting the feature map... The preprocessed low-altitude aerial image is simultaneously fed into two parallel convolutional paths to separate and extract different types of local information. A 1×1 grouped convolution operation is used to extract central feature information, obtaining channel-independent fine-grained information. A 3×3 standard convolution is then used to capture local contextual information to supplement the spatial relationships of the features. Next, the two partial features are concatenated along the channel dimension, followed by convolutional operations for channel compression and feature fusion, resulting in an enhanced upper-branch feature map that highlights the contrast between the central features and the local context. Upper branch feature map The expression can be: In the formula, This indicates a 1×1 grouped convolution operation used to extract central feature information. Here, it represents a standard convolutional block composed of convolution, batch normalization, and ReLU activation function. The superscript indicates... This indicates the use of grouped convolution, with the number of groups being... This refers to the number of channels in the input feature map, where the subscript indicates the shape of the convolution kernel. ; Indicates a channel connection operation; This represents grouped convolution, with the number of groups being... subscript The shape of the convolution kernel; Indicates the input feature map; This represents a standard 3×3 convolution with 1 group. This represents the input feature map.

[0037] For the middle branch 102 of the MFE, the aim is to enhance the model's sensitivity to the directional features of the target contour by utilizing asymmetric strip convolutions through directional modeling. To achieve this goal, the middle branch designs two parallel convolutional paths: vertical-horizontal convolution and horizontal-vertical convolution, to extract complementary feature information from different directions.

[0038] The purpose of vertical and horizontal convolution 1021 is to extract and enhance the contour and boundary features of a target from different directions through cascaded asymmetric convolutions. The specific implementation method is as follows: First, input the feature map... Through a The convolutional group is used for processing to integrate and refine channel information without changing the spatial dimension; then, the feature map is processed by a... Vertical convolution is specifically designed to capture the vertical structure and boundary information of a target; then, through a... Horizontal convolution complements vertical features in the horizontal direction to enhance the representation of lateral details; finally, all directional features are fed into a single convolution. Deep fusion is performed during convolution to generate the final output feature map. Output feature maps obtained based on vertical and horizontal convolutions. The expression can be: In the formula, Indicates that the convolution kernel is convolution, Indicates that the convolution kernel is Horizontal convolution, Indicates that the convolution kernel is Vertical convolution, Indicates that the convolution kernel is convolutional sets, This represents the input feature map.

[0039] Horizontal-vertical convolution and vertical-horizontal convolution are complementary; therefore, horizontal-vertical convolution 1022 aims to extract directional features from opposite directions to capture more comprehensive target boundary information. The specific implementation method is as follows: First, input the feature map... Channel information is integrated through a 1×1 grouped convolution; then, a 3×1 horizontal convolution is used to prioritize capturing the horizontal structural features of the target; next, a cascaded 1×3 vertical convolution supplements the horizontal features in the vertical direction to enhance the coherence of the vertical edges; finally, all extracted features are fed into a 3×3 convolution for deep fusion to generate the final output feature map. Output feature map obtained based on horizontal and vertical convolution. The expression can be: ; Output feature map obtained based on vertical horizontal convolution and horizontal vertical convolution and output feature map By concatenating the two along the channel dimension, the middle branch feature map is obtained. The specific implementation method is as follows: complementary information can be extracted from different directions, and the final feature fusion is performed through a 1×1 convolution. (Middle branch feature map) The expression can be: In the formula, This indicates a convolution with a 1×1 kernel.

[0040] For sub-branch 103 of MFE, the aim is to collaboratively capture the local details and global semantic information of the target through multi-scale context modeling, thereby enhancing the understanding of the target's contextual dependencies. The specific method is as follows: First, using... Grouped convolution preserves the original semantic information within each channel without disrupting the spatial structure; parallel use Dilated convolution is used to effectively perceive the structural features of the target's neighborhood; parallel use is employed. Grouped convolutions directly model global context dependencies. Subsequently, the outputs of the three paths are concatenated and passed through a... The convolutional block undergoes channel compression and depth fusion to obtain the feature map of the next branch. Lower branch feature map The expression can be: In the formula, Indicates that the convolution kernel is convolution, This represents a dilated convolution with an expansion rate of 2, a group number of 1, and a kernel size of 5×5. Indicates that the convolution kernel is Grouped convolution.

[0041] For the feature fusion layer of MFE, , and The data is concatenated along the channel dimension and then uniformly integrated through a final 1×1 convolution to generate a highly discriminative output feature map that combines multi-scale, directional, and central contrast information. (Output Feature Map) The expression can be: ; The following Algorithm 1 is the pseudocode for the Multi-Branch Contextual Feature Fusion (MFE) module. This algorithm extracts multi-scale features from the input features through three parallel branches and performs feature aggregation and semantic enhancement during the fusion stage. The feature map output by the previous stage downsampling convolutional block (CBS) of the input features is shown below. Output feature map This refers to multi-scale enhancement features.

[0042] The multi-branch feature fusion module C2f-MFE aims to deeply optimize the C2f module in the YOLOv8 baseline model. Existing C2f modules can provide rich gradient flows to enhance feature representation through their cross-stage local networks (CSPs), but their internally stacked Bottleneck modules employ homogeneous... Convolutional feature extraction results in a single receptive field, making it difficult to simultaneously capture local details and global contextual information when the target scale varies significantly. To address this, the C2f-MFE design in this embodiment, while maintaining the C2f structure, introduces a multi-branch feature extraction module (MFE) to enhance the representation capability of multi-scale target features by employing multi-branch, multi-scale receptive fields.

[0043] The main function of the feature segmentation unit in the C2f-MFE module is to decouple the input features according to their functions, preserving local gradient information and high-level semantic information respectively. The input to the feature segmentation unit is the feature map output by the upsampled convolutional block (CBS) in the backbone network, denoted as... To balance gradient propagation and deep semantic modeling, this unit employs a channel segmentation strategy to divide the input features... Divide into two parts along the channel dimension, and preserve the gradient features, let them be... Used to preserve gradient flow information; semantic feature extraction, set as This is used for deep feature extraction in subsequent cascaded units. Specific segmentation results. and The expression can be: In the formula, Operation along the channel dimension Divided into two parts, This indicates that the convolution kernel size is It consists of convolutional layers, batch normalization layers, and activation functions in sequence. This represents the segmented feature map. The focus is on preserving spatial details to facilitate gradient flow. Used for subsequent deep semantic feature extraction H represents the number of channels, H represents the height, and W represents the width.

[0044] The main function of the cascaded units of the C2f-MFE module is to extract semantic features from the feature segmentation unit. Multi-level semantic enhancement and feature extraction are performed. Through multiple iterative feature extraction operations, cascaded units create a "semantic accumulation" effect, allowing features to integrate low-level details and high-level abstract information during progressive propagation. This gradually deepens the network's understanding of contextual information, resulting in semantic feature representations with stronger discriminative capabilities. Cascaded units introduce multiple Bottleneck-MFE structural units as basic building blocks to achieve deep modeling of multi-scale contextual features. The iterative processing of cascaded units can be represented as follows: In the formula, It represents the feature extraction operation, and MFE represents the multi-branch feature extraction module. This represents the feature map obtained after the first Bottleneck-MFE processing. This represents the feature map after processing by the i-th unit.

[0045] The main function of the feature fusion unit in the C2f-MFE module is to integrate features from different paths to obtain a highly expressive feature map that combines shallow details with deep semantics. The feature fusion unit includes two processing steps: the first step is feature concatenation, which is responsible for integrating gradient-preserving features from the feature segmentation unit. Semantic feature extraction and deep features from cascaded units The first step is to stitch the features together according to the channel dimension. The second step is feature fusion. The stitched features still have problems such as excessive channel dimension and semantic repetition. Therefore, further channel compression and feature integration are required.

[0046] The calculation formula for feature splicing operation can be: In the formula, This indicates that two shortcut feature maps are collected in the original C2f path. This represents a set of n feature maps generated from stacked Bottleneck-MFE blocks. It's worth noting that all feature maps have the same spatial resolution. And the number of channels is Therefore, the feature dimension after concatenation is .

[0047] The calculation formula for feature fusion operation can be: In the formula, Represents multi-scale enhanced feature maps. Algorithm 2 is the pseudocode for the Multi-Branch Contextual Feature Fusion (C2f-MFE) module. C2f-MFE integrates and enhances multi-scale contextual information through three stages: feature segmentation, cascaded feature extraction, and feature fusion. The input is a feature map. The output is a multi-scale enhanced feature map. This refers to the initial fusion feature map.

[0048] The multi-branch convolutional attention module 11 is connected to the multi-branch feature fusion module and is used to perform channel and spatial dimension attention weighting on the initial fused feature map to obtain the enhanced intermediate feature map.

[0049] To address the interference of complex backgrounds on detection performance in UAV images, Xu et al. proposed YOLOv5s-pp, an improved small target detection model based on YOLOv5s. This method introduces a lightweight Coordinate Attention (CA) module into the backbone network to encode long-range dependencies in the horizontal and vertical directions, thereby highlighting channels relevant to the target and suppressing background responses. Furthermore, YOLOv5s-pp introduces the Meta-ACON activation function and adds an additional small target detection head to enhance the modeling ability of fine-grained small target features while maintaining a relatively compact network structure. However, since CA still relies on global pooling operations, when the background region dominates the image, the generated attention weights may still be biased towards the background rather than the real target. Tahir et al. further addressed the problems of complex backgrounds and occlusion by proposing PVswin-YOLOv8s for UAV pedestrian and vehicle detection tasks. This method replaces the last C2f module in the YOLOv8s backbone network with a stronger global context modeling module to expand the receptive field and enhance long-range dependency modeling capabilities. Simultaneously, it introduces CBAM in the neck network to perform sequential attention modeling on channel and spatial dimensions, thereby strengthening the target region response against complex backgrounds. While this design significantly improves detection accuracy, the introduction of a more complex backbone module and multiple attention mechanisms significantly increases model complexity and computational cost, which is detrimental to real-time UAV applications. Overall, existing research shows that combining long-range contextual information with channel-spatial attention helps suppress interference from complex backgrounds, but there is still a significant trade-off between background suppression capabilities and model lightweighting. To address the problem that existing attention mechanisms are easily dominated by large areas of irrelevant background, thus diluting the response of small target features, this embodiment proposes a multi-branch convolutional attention module (MCAM). Its core lies in guiding attention through multi-scale features, enabling it to accurately focus on the target rather than being diluted by large areas of irrelevant background.

[0050] The multi-branch convolutional attention module is used to suppress background interference in images and enhance weak features of small objects. A schematic diagram of the multi-branch convolutional attention module is shown below. Figure 5 As shown, Figure 5 (a) in the diagram shows the overall structure of MCAM. The input of MCAM is the output feature map of the C2f-MFE backbone network. (b) is the channel attention module diagram, and (c) is the spatial attention module diagram. In (a), the multi-branch convolutional attention module is composed of the feature extraction module, the channel attention module (CA), and the spatial attention module (SA) connected sequentially. The above three modules are in a serial relationship.

[0051] In this embodiment, the MCAM module includes: a multi-branch feature extraction unit, used to process the input initial fusion feature map to generate a basic feature map containing multi-scale contextual information; a channel attention branch, used to generate channel attention weights based on the global information of the basic feature map; a spatial attention branch, used to generate a spatial attention map based on the spatial distribution of the basic feature map; and a feature refinement unit, used to jointly weight the basic feature map using the channel attention weights and the spatial attention map to generate an enhanced intermediate feature map; The implementation process of the multi-branch feature extraction unit 111 of the MCAM module can be referred to the implementation process of the multi-branch feature extraction module (MFE) described above, and will not be repeated here.

[0052] The main function of the channel attention branch 112 of the MCAM module is to model the dependencies between different channels, identify semantic channels that contribute significantly to the detection task, enhance their feature responses, and suppress irrelevant or redundant channels. The input to the channel attention module is a multi-scale enhanced feature map from the output of multiple MFE modules, denoted as... The channel attention module mainly consists of three processing stages. The first stage is global feature extraction, which is used to represent the global semantic response of each channel. The second stage is channel weight generation, which generates channel weights through a compression-activation-restore structure. The third stage is channel weighted fusion, which obtains a weighted feature map.

[0053] For global feature extraction, it is... Global average pooling is performed to compress spatial information into a single-channel description vector, characterizing the semantic response strength of each channel across the entire spatial range, denoted as . Semantic response strength The expression can be: In the formula, Representing feature maps in space The channel vector at that location is the local feature response at that position; Indicates the number of channels is Size is The feature map is defined by H, where H represents the height of the feature map and W represents the width of the feature map.

[0054] Channel weight generation is based on the global semantic description vector. Generate channel weight vector The specific process is as follows: First, a 1×1 convolution is used to compress the channels to reduce dimensionality and model non-linear relationships; then, the ReLU activation function is used to introduce non-linear expressive power; finally, a second 1×1 convolution is used to restore the number of channels, and the Sigmoid function is used to generate the channel weight vector. Channel weight vector The expression can be: In the formula, Indicates the number of channels is Size is Feature map, This represents the SigMoid activation function. express The convolution operation is denoted by ReLU, which represents the activation function.

[0055] Channel weight fusion refers to combining the generated channel weights Acting on input features It can perform channel-wise weighted fusion of enhanced multi-scale feature maps, and the calculation formula is as follows: In the formula, Indicates the number of channels is Size is Feature map, This represents the matrix dot product.

[0056] For the spatial attention branch 113 of the MCAM module, its main function is to enhance the features of the channels. Spatial saliency modeling enables the network to focus on the target region and suppress background interference. The spatial attention module recognizes the varying importance of different spatial locations for the detection task, and the model adaptively assigns spatial weights based on the global context, thereby enhancing the response of target region features. The spatial attention module can be divided into two processing stages: spatial feature aggregation and spatial weighting.

[0057] For spatial feature aggregation, by capturing the global contextual relationships of spatial locations, channel information is compressed into single-channel feature maps to compute the importance response for each spatial location. Specifically, channel dimensionality reduction is achieved through convolution, transforming the input features... Convert to spatial weight mapping Its expression can be: In the formula, This indicates that the number of channels is 1 and the size is [value missing]. Feature map, express Convolution operation, This represents the SigMoid activation function.

[0058] Spatial weighting will yield the feature map Input features Element-wise multiplication is used to weight the saliency of each spatial location, resulting in a weighted spatial weight map, the expression of which can be: In the formula, This indicates that the number of channels is 1 and the size is [value missing]. Spatial attention feature map.

[0059] The results obtained from channel attention and spatial attention are combined again to not only filter and enhance channel information, but also focus and optimize spatial location information, resulting in a two-dimensional enhanced feature map, the expression of which can be: In the formula, Represents a two-dimensional enhanced feature map. Represents spatial attention feature maps. This represents the initial fusion feature map.

[0060] Algorithm 3 is the pseudocode for the Multi-Branch Convolutional Attention Module (MCAM). MCAM achieves adaptive weighting and saliency enhancement of features in two dimensions through a joint mechanism of channel attention and spatial attention. The input is the fused features from the multi-branch contextual feature fusion module. The output is a two-dimensional enhanced feature. Algorithm. First, calculate the channel weights. And generate channel enhancement features Then generate a spatial weight map. And obtain spatial enhancement features Finally, the two are merged to complete the input. Hybrid attention weighting.

[0061] The top-down reverse enhancement module 12 is connected to the multi-branch convolutional attention module, which generates guiding signals based on the deep semantic features contained in the intermediate feature map for weighted regulation and outputs the target detection result.

[0062] To address the issue of small targets, existing research employs feature fusion to reduce feature loss during network propagation. Zhang et al. designed FFM in FFCA-YOLO, optimizing the feature fusion path through channel reweighting to avoid the dilution of target information caused by the step-by-step propagation in traditional FPN. However, while high-level feature maps are semantically rich, they have low spatial resolution, and concatenating them with lower-level features may blur the boundary information of small targets. Li et al. designed SOFA-FPN in EHRPM, utilizing feature maps from the second layer of the backbone network, which contain rich edge information, and combining them with deeper feature maps to enhance the utilization of edge information of small objects. The main drawback of SOFA-FPN is that it gains performance improvement in small target detection by increasing complexity and computational cost. To address the shortcomings of feature fusion methods, such as high-level semantic information blurring the boundaries of shallow small targets and complex fusion strategies increasing computational overhead, this embodiment proposes feature map weighting from the perspective of semantic feedback. This enhances the saliency of small targets by suppressing redundant background information while effectively reducing the computational complexity of the model through channel compression, avoiding the introduction of additional computational burden.

[0063] The Top-Down Reverse Enhancement Module (TDRE) addresses the issue of weak feature responses and susceptibility to background interference for small targets during deep network propagation. TDRE aims to leverage the rich semantic information contained in deep feature maps to guide and optimize shallow feature maps, thereby enhancing the response of small targets in shallow features. While suppressing redundant background information and enhancing the saliency of small targets, channel compression is employed to reduce the computational complexity of the model. A schematic diagram of the Top-Down Reverse Enhancement Module is shown below. Figure 6 As shown, it mainly consists of four parts, the first being the shallow input feature map. The first part contains the target's shape and location information; the second part is the deep feature map. Compared to shallow feature maps It contains stronger semantic information; thirdly, it will Deep feature map after upsampling Fourth, it outputs feature maps by upsampling to enhance detailed information. The result is the inverse weighting of the original shallow features. TDRE consists of two stages: the inverse weight map generation stage and the feature weighting enhancement stage.

[0064] The main function of the reverse weight map generation stage is to establish guiding relationships between shallow features using deep semantic features, generating a weight map that reflects the saliency of spatial semantics. This weight map can indicate regions with strong responses and target significance in deep semantics, providing guidance for subsequent shallow feature enhancement. The specific method is as follows: First, it receives two inputs from both the shallow and deep layers. One input is a shallow, large-scale feature map, denoted as... The other path is a deep, small-scale feature map, denoted as... ;in, and These are the number of channels in the shallow feature map and the number of channels in the deep feature map, respectively, satisfying the condition. ; and These are the heights of the shallow feature map and the deep feature map, respectively; and It is the width of the shallow feature map and the deep feature map, satisfying ,and To enable deep semantics to guide shallow features in the reverse direction, an upsampling operation is employed. The specific method is as follows: Mapped to the same level through upsampling and channel transformation. Using the same spatial scale and normalized by the Sigmoid function to generate an inverse weighted graph. Its expression can be: In the formula, This represents the activation function. express Convolution operation, This indicates an upsampling operation.

[0065] The main function of the feature weighting enhancement stage is to adaptively weight shallow features based on the inverse weight map to enhance the salient regions of small targets and suppress irrelevant background. Therefore, the TDRE module uses inverse weighting... shallow features Weighted adjustments are made to achieve adaptive enhancement of the feature response, and the expression can be: In the formula, This represents the output feature map of the TDRE module. This represents matrix multiplication.

[0066] Algorithm 3 is the pseudocode for the Top-Down Reverse Enhancement Module (TDRE). TDRE achieves semantic guidance and background suppression of shallow features by deep features through a reverse semantic feedback mechanism. The input includes shallow features. with deep features The output is the feature after inverse enhancement. The specific method is as follows: First, an inverse weight map is generated through upsampling and convolution operations. Then use reverse weights Adaptive weighting is applied to shallow features, and the output is the feature after inverse enhancement. .

[0067] Features after inverse enhancement Post-processing the output yields directly understandable target detection results, mainly including: Bounding box decoding: Convert the normalized coordinate offsets predicted by the model into specific bounding box coordinates in the image coordinate system (usually a rectangle that outlines the detected vehicle position).

[0068] Category determination: Based on the category confidence score output by the model, determine the category of the target within each bounding box (such as "small car", "large truck", "motorcycle" etc.).

[0069] Confidence filtering and non-maximum suppression: First, a confidence threshold is set to filter out unreliable detections with excessively low scores. Then, a non-maximum suppression algorithm is applied to merge redundant detection boxes with high overlap, retaining the most accurate bounding box for each target. The final target detection results can be presented in structured data form (such as JSON format, containing the category, confidence score, and bounding box coordinates for each detected target), or the detection results (bounding boxes and category labels) can be visualized and overlaid onto the original image to generate an intuitive detection result map for user or subsequent system analysis.

[0070] Algorithm 5 presents the overall pseudocode flow of MTNet. The input feature is a low-altitude aerial image I, and the output is the algorithm's detection result D. To clearly illustrate the overall processing flow and logical relationships between modules of the MTNet model, this embodiment provides a pseudocode description of the model based on its structural design. This pseudocode aims to depict the entire process from input image to output detection result in a modular manner, highlighting the functional division and information transmission mechanism of each submodule in the feature extraction, fusion, and enhancement stages.

[0071] The MTNet overall detection algorithm mainly consists of five stages. The first is the input preprocessing stage: normalizing and resizing the input aerial images to generate standardized input features, providing basic data for subsequent network processing. The second is the Backbone feature extraction stage: this stage utilizes the proposed C2f-MFE and MCAM modules to enhance semantic information and discriminative capabilities. Specifically, in this stage, the model constructs feature maps at different resolutions (320×320, 160×160, 80×80, 40×40, 20×20) through five downsampling operations, embedding the proposed modules at key levels. Algorithm 2 (C2f-MFE) is called for multi-branch convolutional feature extraction to enhance the multi-scale representation of features; Algorithm 3 (MCAM) is called to apply attention weighting to the extracted features, improving the representation of salient targets; C2f-MFE and MCAM are used together to form the first output of the Backbone. (80×80) is used for subsequent detection of small targets; C2f-MFE + MCAM is applied again to obtain the second output of the Backbone. (40×40) is used for detecting medium targets; C2f-MFE and SPPF are called to perform multi-scale pooling on a 20×20 scale to generate the third output of the Backbone. The algorithm is used to detect large targets. The third stage is the Neck feature fusion stage: this stage aims to achieve semantic fusion and contextual enhancement of multi-layer features through upsampling, concatenation, and inverse enhancement operations. When calling Algorithm 4 (TDRE), shallow features are enhanced through inverse attention weighting, suppressing background noise and enhancing the feature expression of small targets. The fourth stage is the Head detection and prediction stage: detection is performed based on feature maps of different sizes, and finally, target classification and bounding box regression are performed by multi-layer detection heads. The fifth stage is the result output stage: integrating the prediction results of each detection layer, outputting the final target category and location coordinates, realizing multi-scale small target detection. Algorithm 5: The pseudocode of MTNet, its input: aerial image I; output: detection result D, as follows: This invention proposes a low-altitude aerial vehicle target detection model, MTNet, to improve the accuracy of vehicle target detection in UAV scenarios. The three main contributions are as follows: First, C2f-MFE is proposed, which uses a multi-branch convolutional structure to extract multi-scale feature information for targets of different scales in different aerial images, enhancing feature richness and improving the expressive power of the final output features in multiple dimensions including spatial, semantic, and receptive field. Second, MCAM is proposed, addressing the insufficient target feature extraction caused by complex background information. It uses MFE for multi-dimensional feature extraction, followed by a channel attention mechanism to strengthen target features, with spatial attention focusing on salient regions. Third, TDRE is proposed, addressing the weak features of small targets. It utilizes deep semantic information to suppress redundant responses in shallow features by generating a reverse attention mask in the deep layers of the network, weighting shallow features, suppressing background while enhancing the weak features of small targets.

[0072] To verify the target detection effect of the present invention, the implementation process of the present invention was experimentally verified, including: 1. Configure the experimental environment.

[0073] The experiments of this invention were conducted using Python 3.8 and PyTorch 2.0.0 frameworks, accelerated by CUDA 11.8 and cuDNN, and ran on Ubuntu 20.04 operating system. The experimental platform was equipped with a 12-core Intel(R) Xeon(R) Platinum8352V CPU 2.10GHz and an NVIDIA GeForce RTX 4090D graphics card with 24GB of video memory. During the model training phase, stochastic gradient descent (SGD) was used as the model optimizer, with an initial learning rate (LR) set to 0.01, a total of 200 training epochs, and a batch size of 10. It should be noted that the model proposed in this embodiment did not use any pre-trained weights during the entire training process.

[0074] 2. Obtain the dataset for model training.

[0075] All experiments in this invention were conducted on two typical drone aerial vehicle detection datasets: DroneVehicle and VEDAI.

[0076] (1) The DroneVehicle dataset is a large-scale benchmark dataset specifically designed for vehicle detection tasks from the perspective of drones. It contains 28,439 images, of which 17,990 are for training, 8,980 for testing, and 1,469 for validation. The data covers a variety of complex scenes, including urban roads, rural areas, residential areas, and parking lots. The images were captured both day and night, and the resolution is 640×512. This dataset features a typical drone overhead view, significant target scale variations, and diverse backgrounds, which enhances the challenge of the detection task.

[0077] (2) VEDAI is an aerial imagery dataset for vehicle detection. This dataset contains cropped sub-images from aerial images taken by the Automated Geo-Reference Center (AGRC) in Utah. The original AGRC images were taken at the same altitude with a resolution of 12.5 cm per pixel. VEDAI manually selected and cropped 1210 images of size 512×512 from these images, with 1090, 120, and 120 images used for training, testing, and validation, respectively. The dataset covers various backgrounds, including farmland, grassland, mountains, cities, and roads. The task is to detect eight types of vehicles.

[0078] 3. Determine the evaluation indicators for the detection effect.

[0079] To comprehensively evaluate the performance of the proposed MTNet model in small object detection tasks, the model is quantitatively analyzed from two aspects: accuracy and performance testing.

[0080] (1) Accuracy metrics are used to measure the accuracy and recall of a model at the level of detection results, and are the core evaluation criteria for target detection tasks.

[0081] (a) Precision refers to the proportion of predictions that correctly identify the positive class that actually are positive. A higher precision indicates higher accuracy, meaning fewer false positives. The formula for precision is: In the formula, TP represents a true positive and FP represents a false positive.

[0082] (b) Recall refers to the number of instances correctly predicted as positive by the model out of all true positive instances. A higher recall value means fewer missed detections by the model. The formula for calculating recall is: ; where TP represents the true negative and FN represents the false negative.

[0083] (c) The average precision (AP) is obtained by calculating the area under the Precision-Recall (PR) curve. A higher AP value indicates better overall model performance. The formula for AP is: ; (d) mAP0.5 is the mean precision, representing the average AP value across all classes in the dataset. This is the most crucial comprehensive evaluation metric, fully reflecting the overall detection performance of the model. A higher mAP value indicates better overall model performance. For N classes, the formula for calculating mAP is: In the formula, N is the number of categories. It is the average precision of the i-th category when the IOU threshold is 0.5.

[0084] (e) mAP0.5:0.95 is the average mAP across 10 IoU thresholds (from 0.5 to 0.95, with a step size of 0.05). It is more stringent than mAP0.5, requiring the model to perform well across multiple IoU thresholds, and can be expressed as: In the formula, It is the average precision calculated for all categories when the IOU threshold is t.

[0085] (2) Performance testing indicators mainly measure the computational efficiency and resource utilization of the model during the inference phase.

[0086] (a) The parameter quantity index refers to the learnable parameters that need to be optimized during the training of convolutional layers, fully connected layers, and bias terms in a neural network, and can be expressed as: In the formula, L represents the total number of layers in the network. , , , These represent the number of input channels in the i-th layer, the height and width of the feature map, and the size of the convolution kernel, respectively.

[0087] (b) The computational complexity metric GFLOPs represents billions of floating-point operations per second, which can be expressed as: In the formula, N represents the batch size, and H and W represent the height and width of the input feature map. and These represent the number of input and output channels, respectively. and These represent the height and width of the convolution kernel, respectively, and 2 indicates that each convolution operation requires two floating-point operations.

[0088] This invention employs a series of experimental systems to verify the improvements of the proposed MTNet model in terms of small target detection accuracy, background suppression capability, and feature representation. The experiments include comparative experiments, ablation experiments, and visualization experiments.

[0089] First, comparative experiments were conducted to verify the overall superior performance of the model. To verify the performance of MTNet in low-altitude aerial vehicle small target detection, this paper compared the performance of the MTNet model with several baseline models on DroneVehicle and VEDAI benchmarks. The baseline models included mainstream detectors from the YOLO series, mainstream two-stage target detectors, and representative models optimized for aerial photography in recent years such as BRSTD, Drone-YOLO, and EHRPM.

[0090] Second, ablation experiments were conducted to evaluate the performance of MTNet and analyze the independent contributions and synergistic effects of key MTNet modules, including C2f-MFE, MCAM, and TDRE. Two sub-experiments were set up for this purpose.

[0091] (1) Ablation experiments of each component. In order to fully evaluate the performance of MTNet, the independent contributions and synergistic effects of the core components C2f-MFE, MCAM and TDRE were analyzed, and ablation experiments were carried out on the DroneVehicle dataset.

[0092] (2) MCAM module ablation experiment. To further verify the effectiveness of the proposed multi-branch convolutional attention module (MCAM), this paper uses mainstream attention mechanism modules as baseline comparison models. Specifically, these include channel spatial attention mechanism (CBAM), efficient channel attention mechanism (ECA), global attention (GAM), and coordinate attention (CoordAtt). The best results are highlighted in bold.

[0093] Third, visualization experiments are used to demonstrate the intuitive improvements of the model in feature extraction and target recognition. Two sub-experiments are set up here.

[0094] (1) Visualization experiment of MTNet and other comparison algorithms. In order to verify the detection performance of the proposed model, this study selected images of different scenes and used Ground Truth annotation as a reference to compare and analyze the detection results of MTNet with YOLOv8n, BRSTD, Drone-YOLO and EHRPM.

[0095] (2) Visualization experiment of TDRE module. To verify the technical effect of TDRE module in feature representation and semantic reverse enhancement in this invention, a visualization comparison experiment was conducted. The experiment compared heatmaps at different stages.

[0096] 4. Ablation experiment.

[0097] (1) Overall ablation experiment The ablation experiment results of MTNet on the DroneVehicle dataset are shown in Table 1: Table 1 The above experiments use MTNet as the primary model and variants of MTNet as comparative models. The C2f-MFE module is denoted as A, the MCAM module as B, and the TDRE module as C. For example, model MTNet-A represents an MTNet model that does not use C2f-MFE. √ indicates the module used for the corresponding column, × indicates the module not used for that column, and the optimal results are all bolded.

[0098] As shown in Table 1, the detection performance of the MTNet model continuously improves after the gradual introduction of each module, demonstrating the effectiveness and complementarity of the proposed structural design. Firstly, overall performance is improved. Compared with the baseline model YOLOv8n, the MTNet model improves mAP0.5 by 3.6%, mAP0.5:0.95 by 3.3%, P by 4.4%, and R by 2.5%, ultimately achieving an mAP0.5 of 74.1%. With only a small increase in parameters (3.32M), the model achieves an ideal balance between detection accuracy and computational efficiency, fully validating the effectiveness and overall optimization advantages of the proposed structural design. Secondly, the contributions of each module are significant. Compared to the baseline model, MTNet-BC improves mAP0.5 by 1.1% and mAP0.5:0.95 by 1.0%, with only a 0.26M increase in parameters, indicating that this module can effectively improve detection accuracy with a relatively small computational cost. After introducing the MCAM module, MTNet-AC achieved a 1.5% improvement in mAP0.5 and a 1.4% improvement in mAP0.5:0.95, making it the single-module combination with the most significant performance improvement and validating the effectiveness of this mechanism in complex scenarios. Compared to the baseline model, MTNet-AB achieved a 1.3% improvement in mAP0.5 after introducing TDRE, while reducing the number of parameters by 0.62M. TDRE not only possesses inverse feature enhancement capabilities but also achieves model lightweighting through channel compression, validating that this module combines feature enhancement and efficiency. Thirdly, there is a synergistic effect between modules. MTNet-B achieved a 1.8% improvement in mAP0.5 and reduced the number of parameters to 2.84M, achieving a dual optimization of performance and complexity; MTNet-A outperformed all single-module and dual-module combinations, with a 2.4% improvement in mAP0.5, demonstrating a strong synergistic effect between the attention mechanism and the inverse enhancement strategy. Ultimately, the MTNet model integrating all three modules achieved the best results on all four core metrics, indicating that the C2f-MFE, MCAM, and TDRE modules provide effective solutions to the three key problems of multi-scale features, complex backgrounds, and feature attenuation of small targets, respectively. The three modules work together to build an efficient and robust detection framework.

[0099] (2) MCAM module ablation experiment.

[0100] The comparison results between different attention mechanisms are shown in Table 2: Table 2 The above experiments use MTNet as the main description. The multi-branch convolutional attention mechanism (MCAM) of this invention is denoted as A, the channel spatial attention mechanism (CBAM) as B, the efficient channel attention mechanism (ECA) as C, the global attention mechanism (GAM) as D, and the coordinate attention mechanism (CoordAtt) as E. The model MTNet-A+B indicates that channel spatial attention (CBAM) was used instead of multi-branch convolutional attention (MCAM), and the rest are similar.

[0101] As shown in Table 2, MTNet achieved the best overall detection performance among different attention mechanisms. First, it demonstrated the best overall performance. MTNet achieved the best overall detection accuracy, with mAP0.5 and mAP0.5:0.95 reaching 72.0% and 50.0% respectively, a 1.1 percentage point improvement over MTNet-A+B based on CBAM, validating the significant advantage of the multi-branch structure of the MCAM module in feature enhancement. Second, it demonstrated stronger fine-grained recognition capabilities. In the more challenging Truck and Freight categories, MTNet achieved accuracies of 70.7% and 49.3% respectively, making it the only model among all comparison models to achieve the best performance in both categories simultaneously. This indicates that the MCAM module can effectively improve the fine-grained discrimination capability of features in aerial photography scenarios with complex backgrounds and significant differences in category scale. Third, in terms of computational efficiency, MTNet achieved the most significant accuracy improvement with moderate computational overhead. Its computational cost is slightly higher than that of MTNet-A+E using the CoordAtt module (GFLOPs are 12.1 and 8.1 respectively), but the improvement in mAP of 0.5:0.95 reaches 1.7%, indicating that the multi-branch attention design of the present invention achieves a superior balance between accuracy and complexity.

[0102] 5. Comparative experiment.

[0103] (1) Comparison of MTNet and existing methods on the VEDAI dataset.

[0104] Table 3 shows the comparison of the mean accuracy (mAP 0.5) of MTNet and existing methods on the VEDAI dataset. Compared to the DroneVehicle dataset, VEDAI introduces a richer variety of target categories, increasing the diversity and challenge of the detection task.

[0105] Table 3 As shown in Table 3, firstly, in terms of overall performance, MTNet demonstrates the best overall detection performance, achieving an mAP of 0.5 of 69.1%, ranking first among all compared methods. Compared to the baseline model YOLOv8n (63.7%), MTNet achieves a significant improvement of 5.4 percentage points and is significantly better than the advanced method BRSTD (65.5%) optimized for aerial photography tasks. This indicates that the modular structure proposed in this invention not only works effectively in specific scenarios but also possesses strong cross-scenario generalization capabilities. Secondly, in terms of category-level performance, MTNet shows particularly significant improvements in identifying the more difficult Other (67.2%) and Tractor (70.3%) categories, increasing by 9.6% and 14.9% respectively compared to the baseline model. This demonstrates that the structural design of this invention can more effectively capture small target information in complex backgrounds and improve the detection capability for target categories with weak features and large scale differences.

[0106] (2) Comparison of MTNet and existing methods on the DroneVehicle dataset.

[0107] Table 4 shows the performance comparison results of MTNet and existing methods on DroneVehicle: Table 4 As shown in Table 4, MTNet achieves a significant advantage in core detection accuracy. Firstly, in terms of overall performance, MTNet achieves an mAP of 74.1%, ranking first among all compared methods, a 3.6 percentage point improvement over the baseline model YOLOv8n, and a 2.1 percentage point improvement over recent advanced algorithms like EHRPM. Secondly, in terms of model efficiency, MTNet also demonstrates superior lightweight characteristics. As a single-stage detector, its parameter count is only 3.32 M and its computational cost is 17.0 GFLOPs, far lower than two-stage methods such as Faster R-CNN (41.15 M, 211.28 GFLOPs). Compared to other lightweight models (such as YOLOv5n and YOLOv6n), MTNet achieves an accuracy improvement of over 5% with extremely high parameter utilization efficiency, achieving an ideal balance between accuracy and computational complexity. Thirdly, in terms of category-level performance, MTNet also excels in recognizing difficult-to-detect categories. In categories such as Truck (72.7%) and Freight (51.2%), where the target scale varies greatly and is easily confused with the background, MTNet's detection accuracy improved by more than 4.5 percentage points compared to the baseline, effectively verifying the relevance and effectiveness of the proposed module in dealing with complex background problems.

[0108] 6. Visualization experiments.

[0109] The thermal effect diagram of TDRE is as follows Figure 7 As shown. Figure 7 In the heatmap, brighter areas indicate stronger feature responses at that location, suggesting the model's attention is more focused on that area. To verify the effectiveness of this module, we compared heatmap representations at different levels. Figure 7 (a) in the image is the original image, which presents a city street scene from an overhead view, including elements such as roads, vehicles and buildings, and is the basis for subsequent feature extraction. Figure 7 (b) is a shallow heatmap, which shows a relatively diffuse thermal region. Although it can roughly outline the contours and distribution of objects, it also contains a lot of background noise, and the feature response of small targets is very weak. Figure 7 (c) in the figure is a deep heat map, which shows a more concentrated heat area, indicating that the model has extracted higher-level semantic information. However, due to the low resolution of the deep feature map, the localization is not accurate enough, and there is still redundant attention to non-target areas. Figure 7 (d) in the figure is a simple spliced ​​heat map. As can be seen from the figure, the heat area is very diffuse. Strong background noise from the shallow layer and vague semantic information from the deep layer are mixed together equally, which leads to serious distraction. Figure 7 In Figure (e), the heatmap after TDRE processing contrasts sharply with the diffuse state of heatmap (d). In (d), the background area is almost completely darkened, while the area representing the vehicle target appears as bright and highly concentrated heat points. Therefore, the visual changes clearly demonstrate that the TDRE module has achieved significant results.

[0110] The visualization comparison results of MTNet object detection on the DroneVehicle dataset are shown in the figure below. Figure 8 As shown. Figure 8 As shown in (a), in densely trafficked scenarios, MTNet exhibits superior recognition completeness, effectively reducing the false negatives commonly found in other models; Figure 8 As shown in (b), in complex scenes where the background and target features are similar, MTNet exhibits stronger robustness and significantly reduces false positives and false negatives thanks to its superior background suppression capabilities; Figure 8 As shown in (c), under harsh conditions such as low light, MTNet maintains the highest detection accuracy through its optimized feature extraction structure, while other models suffer from false detections, missed detections, and class confusion; furthermore, as Figure 8 As shown in (d) in the figure, in multi-scale target scenes, MTNet not only accurately identifies targets of different sizes, but also successfully detects unlabeled objects in the dataset, demonstrating its powerful multi-scale modeling ability and generalization potential.

[0111] In summary, the visual comparison clearly demonstrates that MTNet outperforms the comparison models in areas such as dense small target recognition, complex background suppression, low-light robustness, and multi-scale target detection, showcasing its excellent feature learning capabilities and practical application potential.

[0112] This invention provides a deep learning-based method for detecting small targets of vehicles in low-altitude aerial photography. By introducing a C2f-MFE multi-scale feature fusion module, an MCAM multi-branch convolutional attention module, and a TDRE top-down inverse enhancement module, a complete optimization chain is formed in three aspects: multi-scale feature extraction, feature focusing, and feature enhancement. Each module plays a role in addressing the main difficulties of small target detection in aerial photography, and experimental results demonstrate significant performance improvements.

[0113] First, let's discuss the effects and functions of the C2f-MFE module. The C2f-MFE module primarily enhances the network's multi-scale feature representation capabilities to address the issue of large target scale variations in aerial images. This module achieves feature segmentation, cascaded extraction, and feature fusion through a multi-branch structure, allowing semantic information from different scales to fully interact in a unified feature space, thereby improving the model's ability to identify small targets. On the DroneVehicle dataset, introducing only the C2f-MFE module improved mAP0.5 by 1.1% and mAP0.5:0.95 by 1.0%; on the VEDAI dataset, the model's detection accuracy improved by 1.3%. These results demonstrate that C2f-MFE effectively alleviates the information loss problem caused by scale inconsistency, providing the model with a more hierarchical feature representation.

[0114] Secondly, regarding the effects and functions of the MCAM module, it addresses the issue of complex backgrounds in aerial photography by introducing a dual-dimensional attention mechanism involving both channel and spatial dimensions. Through a multi-branch convolutional structure, MCAM can select discriminative features at the channel level while focusing on the target region at the spatial level, thereby suppressing interference from non-target regions. Ablation experiments show that when the MCAM module is introduced alone, mAP0.5 is improved by 1.5%, mAP0.5:0.95 is improved by 1.4%, and recall is improved by 1.2%; in complex scenes, the false positive rate is reduced by approximately 12%. This indicates that MCAM effectively improves the model's feature selectivity in complex backgrounds, enabling the model to more accurately capture the salient feature regions of small targets.

[0115] Next, we examine the effects and functions of the TDRE module. Addressing the issue of feature attenuation in small targets during network transmission, the TDRE module introduces a top-down inverse enhancement mechanism. This mechanism utilizes deep semantic features to provide semantic feedback to shallow features, guiding shallow feature enhancement by generating an inverse weight map. This allows the model to strengthen weak signals while maintaining structural integrity. Experimental results show that adding the TDRE module improves mAP0.5 by 1.3%, mAP0.5:0.95 by 1.0%, and recall by 1.1%. Simultaneously, the model parameter count decreases by 0.62M, indicating that TDRE not only improves feature representation quality but also enhances the model's lightweight nature. Visualization analysis further demonstrates that the TDRE module effectively strengthens the edge response of small targets and suppresses background noise, resulting in clearer feature focusing.

[0116] Finally, regarding the overall performance of the three modules working together, MTNet achieves end-to-end optimization of feature extraction, attention enhancement, and semantic feedback when the three modules work in tandem, significantly outperforming the baseline model. On the DroneVehicle dataset, the model achieves an mAP of 74.1% and an mAP of 51.9% for mAP 0.5:0.95; on the VEDAI dataset, the mAP of 0.5 reaches 69.1%. Compared to YOLOv8n, the mAP of 0.5 is improved by 3.6% and 5.4% respectively, while the number of parameters only increases by about 7%. This demonstrates that the present invention achieves an ideal balance between lightweight design and high-accuracy detection through efficient modular design, possessing application potential in scenarios such as drones and low-power embedded platforms.

[0117] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for detecting small targets of vehicles in low-altitude aerial photography based on deep learning, characterized in that, The method includes the following steps: A low-altitude aerial image containing the vehicle is acquired, and the image is input into a detection model for processing to output the target detection result. The processing steps are implemented sequentially through the following modules: The multi-branch feature fusion module is used to extract and fuse contextual information and fine-grained information from different receptive fields based on the low-altitude aerial image to obtain an initial fused feature map. A multi-branch convolutional attention module, connected to the multi-branch feature fusion module, is used to perform channel and spatial dimension attention weighting on the initial fused feature map to obtain an enhanced intermediate feature map; The top-down reverse enhancement module is connected to the multi-branch convolutional attention module. It is used to generate guiding signals based on the deep semantic features contained in the intermediate feature map, perform weighted regulation, and output the target detection result.

2. The method for detecting small targets of vehicles in low-altitude aerial photography based on deep learning according to claim 1, characterized in that, The process of obtaining the initial fusion feature map includes: The multi-branch feature fusion module includes multiple convolutional branches set in parallel. Each convolutional branch has a receptive field of different size and is configured to perform directional modeling to extract local contextual features, global contextual features, and fine-grained directional information representing target edges and contours from the low-altitude aerial image, respectively. The multi-branch feature fusion module is further configured to fuse the local context features, global context features, and fine-grained directional information extracted from each of the convolutional branches to generate the initial fused feature map.

3. The method for detecting small targets of vehicles in low-altitude aerial photography based on deep learning according to claim 2, characterized in that, The convolutional branch includes: The upper branch is used to enhance the model's ability to distinguish between the core target area and the local background through center-surround contrast modeling; The middle branch is used to enhance the model's sensitivity to the directional features of the target contour by utilizing asymmetric strip convolutions through directional modeling. The next branch is used to collaboratively capture the local details and global semantic information of the target through multi-scale context modeling, so as to enhance the understanding of the target's contextual dependencies.

4. The method for detecting small targets of vehicles in low-altitude aerial photography based on deep learning according to claim 2, characterized in that, The multi-branch feature fusion module also includes: The feature segmentation unit is used to decouple the input features according to their functions, retaining local gradient information and high-level semantic information respectively; Cascaded units are used to perform multi-level semantic enhancement and feature extraction on the semantic features extracted by the feature segmentation units; The feature fusion unit is used to integrate features from different paths to obtain the initial fused feature map, which contains both shallow details and deep semantics.

5. The method for detecting small targets of vehicles in low-altitude aerial photography based on deep learning according to claim 1, characterized in that, The multi-branch convolutional attention module includes: A multi-branch feature extraction unit is used to process the input initial fused feature map to generate a basic feature map containing multi-scale contextual information. The channel attention branch is used to generate channel attention weights based on the global information of the base feature map; A spatial attention branch is used to generate a spatial attention map based on the spatial distribution of the base feature map. And a feature refinement unit, used to jointly weight the basic feature map using the channel attention weights and the spatial attention map to generate the enhanced intermediate feature map.

6. The method for detecting small targets of vehicles in low-altitude aerial photography based on deep learning according to claim 1, characterized in that, The output of the target detection result includes: A reverse attention mask is generated based on the deep semantic features, serving as the guiding signal; The intermediate feature map is weighted using the guiding signal to obtain a weighted feature map. The weighted feature map is subjected to channel compression processing, and the target detection result is determined based on the result of the channel compression processing.

7. The method for detecting small targets of vehicles in low-altitude aerial photography based on deep learning according to claim 1, characterized in that, The overall architecture of the detection model is YOLOv8.

8. The method for detecting small targets of vehicles in low-altitude aerial photography based on deep learning according to claim 1, characterized in that, The detection model is deployed on an embedded platform of an unmanned aerial vehicle (UAV).

Images