Unmanned aerial vehicle aerial image target detection method based on spatial detail perception enhancement and cross-layer semantic fusion

Abstract

The application discloses a kind of based on spatial detail perception enhancement and cross-layer semantic fusion UAV aerial image target detection method, the method includes: the pretreatment of input UAV image or video frame;Through the preset light multi-scale feature extraction network, multi-scale feature extraction is carried out to input image;Through the preset local feature information enhancement module, local information is enhanced to multi-scale feature, and enhanced feature is obtained;Through the preset multi-scale feature aggregation network, multi-scale semantic information fusion is carried out to enhanced multi-scale feature, and fusion feature is obtained;Through the preset spatial detail perception enhancement module, spatial detail is enhanced to fusion feature;After the feature of enhanced spatial detail is input to the preset three groups of detection head, detection result is obtained.The method effectively improves the perception ability of model to different scale targets of UAV aerial image, and improves the detection precision of small target.

Description

TECHNICAL FIELD

[0001] The application relates to a UAV aerial image target detection method based on spatial detail perception enhancement and cross-layer semantic fusion, and belongs to the field of computer vision. BACKGROUND

[0002] In recent years, remote sensing image target detection technology plays an increasingly important role in many key fields such as traffic flow monitoring, disaster response evaluation and military reconnaissance. Remote sensing images are mainly obtained from space or air by satellite or unmanned aerial vehicle platforms. Compared with satellite platforms, unmanned aerial vehicles have the advantages of low manufacturing cost, flexible deployment and moderate flight height, and are particularly outstanding in obtaining medium and low altitude high resolution remote sensing data, and have become one of the core tools for air-to-ground perception.

[0003] Based on the characteristics of the platform, the target detection technology based on unmanned aerial vehicle aerial images realizes the rapid positioning and identification of ground multi-class targets. However, due to the low operating height of the unmanned aerial vehicle and the wide range of shooting, the targets in the images obtained by the unmanned aerial vehicle are often small in size, densely distributed, severely occluded and complex in background. These characteristics make the target detection task more susceptible to environmental interference, resulting in false detection or missed detection. Therefore, how to realize efficient and accurate detection of multi-target, especially small targets in unmanned aerial vehicle aerial images, has become an important research topic in this field.

[0004] With the development of deep learning technology, researchers have proposed a variety of general target detection frameworks, which can be mainly divided into two-stage detection methods (such as Faster R-CNN series algorithms) and one-stage detection methods (such as SSD, YOLO series, etc.). Among them, the two-stage method has higher detection accuracy through the decoupling mechanism of candidate region generation and fine classification regression, but has large computational overhead and slow inference speed, which is difficult to meet the real-time requirements of the unmanned aerial vehicle platform. In contrast, the one-stage method has the advantages of simple structure, fast speed and convenient deployment, and is more suitable for application in resource-constrained unmanned aerial vehicle environments, especially in online small target detection.

[0005] Although existing detectors have achieved excellent performance on conventional natural image datasets, they still face the following key challenges in the task of small target detection in unmanned aerial vehicle aerial images:

[0006] Small target features are easily lost. Since small targets account for a very small proportion in the original image, their features are easily compressed or eliminated after multiple downsampling by deep networks, making it difficult for the model to learn effective semantic representations;

[0007] Multi-scale and dense distribution. The target categories are diverse, the scale difference is large, and the spatial density is high, requiring the model to have good scale adaptability and spatial perception ability;

[0008] Complex backgrounds can cause strong interference. Complex surface information such as buildings, roads, and woodlands can easily introduce background noise, affecting the model's ability to identify target boundaries and textures, and thus exacerbating detection errors.

[0009] To address these issues, academia and industry have proposed numerous improvement strategies. For example, Feature Pyramid Network (FPN) and its variants enhance shallow features through top-down semantic propagation, improving the detection capability of multi-scale targets; lightweight models (such as YOLOv5n, YOLOv8s, YOLOv11s, and LMANet-s) balance speed and accuracy while being more suitable for real-time deployment on edge devices; Transformer-based end-to-end detection models (RT-DETR) have high accuracy, but they lack sufficient preservation of feature details, global features easily replace local features, and require significant computational resources.

[0010] These methods still have many limitations when dealing with target detection in complex backgrounds in UAV images, mainly in terms of insufficient spatial detail modeling capabilities, inadequate multi-scale feature fusion, and insufficient utilization of boundary and structural information. Therefore, there is an urgent need for a UAV aerial image target detection method that can enhance spatial detail perception capabilities, improve feature fusion efficiency, and has good scale adaptability to address the increasingly severe detection challenges in low- and mid-altitude remote sensing scenarios. Summary of the Invention

[0011] To address the problems existing in existing methods, the present invention aims to provide a target detection method for UAV aerial images based on spatial detail perception enhancement and cross-layer semantic fusion.

[0012] To achieve the above objectives, the present invention adopts the following technical solution:

[0013] Step 1: Data acquisition and preprocessing. Obtain the drone aerial image dataset and preprocess it.

[0014] Step 2: Construct the object detection model;

[0015] In the feature extraction stage, a pre-defined multi-scale feature extraction network is used, unlike traditional algorithms that extract three sets of high-level features. This method extracts four sets of feature maps A at different scales from high to low levels. iThe feature maps are defined as follows: a bottom-level feature map A1, a middle-level feature map A2, a high-level feature map A3, and a deep-level feature map A4. Feature maps A1, A2, and A3 are enhanced with local feature information through a preset local feature information enhancement module. The local feature information enhancement module includes a parallel multi-branch structure of standard convolutional paths and multi-scale dilated convolutional paths. It fuses local information under different receptive fields through residual connections and enhances the expressive ability of small target edge structures and contextual information, generating three sets of locally enhanced feature maps B1, B2, and B3.

[0016] In the feature fusion stage, three sets of fused feature maps C are obtained through a pre-set multi-scale feature fusion network. i The three sets of locally enhanced feature maps B i Using a deep feature map A4 as input, this system innovatively constructs three bottom-up upsampling paths and two lateral feature transfer paths for multi-scale feature fusion. A pre-defined lightweight bidirectional attention feature aggregation module replaces the traditional upsampling and concatenation structure to enhance the expressive power of the fused features. A pre-defined spatial detail perception enhancement module enhances the model's ability to focus on spatial location and edge texture in the feature map, enabling more accurate perception of small targets. The lightweight bidirectional attention feature aggregation module consists of an attention mechanism and a bidirectional feature fusion network with dual feature inputs. The spatial detail perception enhancement module is a three-branch heterogeneous structure composed of depthwise separable convolutions, dilated convolutions, group convolutions, variable convolutions, and an attention mechanism, obtaining three sets of spatially enhanced feature maps D. i ;

[0017] In the prediction phase, unlike the common practice of using three sets of detector heads that receive the original input image at resolutions of 1 / 8, 1 / 16, and 1 / 32, this innovative approach removes the detector head that receives the original input image at 1 / 32 resolution and introduces a detector head that receives the original input image at 1 / 4 resolution to better detect smaller targets. Ultimately, the three sets of detector heads (1 / 4, 1 / 8, and 1 / 16) receive three sets of spatially detailed feature maps D respectively. i Construct a loss function to better address the characteristics of small targets being small in scale and densely packed.

[0018] Step 3: Train the model using the preprocessed drone aerial image dataset, and test the best trained model on the test set to obtain the test results.

[0019] Furthermore, in step 1, the VicDrone aerial image dataset is cleaned by deleting files with incorrect labels or corrupted images.

[0020] Furthermore, in step 2, a target detection model for UAV aerial images is established: during the feature extraction stage, a multi-scale feature extraction network is used to extract features from the input image at multiple scales to improve the model's global perception capability and obtain four sets of feature maps A at different scales. i For three sets of feature maps at different scales, A i Feature enhancement is performed using a pre-defined local feature information enhancement module to strengthen the local feature information carried by feature maps at different scales, resulting in three sets of enhanced local information feature maps B at different scales. i Feature map B is enhanced by using a pre-defined multi-scale feature aggregation network to enhance local information at three different scales. i Multi-scale feature aggregation is performed on the original feature map A4 at the smallest scale to obtain three sets of aggregated feature maps C at different scales. i Three sets of aggregated feature maps at different scales C i After passing through the preset spatial detail perception enhancement module, three sets of enhanced spatial detail feature maps D are obtained. i The three sets of enhanced spatial detail feature maps D i The input is fed into three preset detection heads for multi-target detection. The scales of the input feature maps corresponding to the detection heads are 1 / 4, 1 / 8 and 1 / 16 of the original input image, respectively, which correspond to the detection of small, medium and large targets.

[0021] Furthermore, the local feature information enhancement module aims to improve the spatial detail representation and contextual semantic awareness of the target region, thereby enhancing the recognition effect of small targets in complex backgrounds. This module consists of multiple parallel branches, each processing features at different scales and directions to achieve multi-dimensional information fusion. The input features are first compressed through multiple parallel 1×1 convolutional channels, and then fed into dilated convolutional paths containing standard convolutions and different dilation rates (e.g., 1, 3, 5) to expand the receptive field and enhance the ability to model the context. The feature maps output from each branch are concatenated and fused along the channel dimension, and then dimensionality restoration and feature integration are performed through 1×1 convolutions. Finally, the fused features are added to the initial input features through residual connections to preserve the original information and enhance detail responses, effectively mitigating the gradient vanishing problem and improving the stability of model training and detection performance.

[0022] Furthermore, the multi-scale feature aggregation network employs bottom-up and top-down feature transfer mechanisms to achieve efficient information fusion. Specifically, the network receives three sets of local enhanced feature maps B. iA set of original feature maps A4 at the smallest scale is used, and through layer-by-layer lateral connections and weighted fusion operations, cross-layer information interaction is performed on feature maps of different scales, strengthening the transmission of semantic information in shallow layers and compensating for detailed information in deep layers. During the fusion process, upsampling and downsampling operations are introduced to align the scale of feature maps of different resolutions, ensuring the consistency and complementarity of features. Finally, three sets of aggregated feature maps C are output. i Their spatial resolutions are 1 / 4, 1 / 8, and 1 / 16 of the input image, respectively, and are used for small, medium, and large target detection tasks.

[0023] Furthermore, the spatial detail-aware enhancement module aims to further enhance the spatial structure information and local detail representation capabilities of the aggregated feature map. The module first performs depthwise separable convolution processing on the input feature map to extract preliminary spatial information, then branches it into three heterogeneous enhancement paths: the first path uses 1×1 convolution for dimensionality reduction and then connects to an ACBS unit composed of dilated convolution and BN+SiLU to strengthen long-range spatial dependencies; the second path uses 1×1 convolution for dimensionality reduction and then enters a grouped convolution module to achieve efficient local structure modeling; the third path uses 1×1 convolution for dimensionality reduction and then introduces deformable convolution to adapt to target shape and geometric distortion. The feature maps output from the three paths are concatenated along the channel dimension and then fed into the SimAM attention mechanism for joint attention weighted fusion of spatial and channel dimensions, ultimately outputting a detail-aware enhanced feature map D. i This further enhances the model's ability to detect small and edge targets in complex scenarios.

[0024] Furthermore, the three preset detection heads detect three sets of targets corresponding to different resolutions, with corresponding spatial resolutions of 1 / 4, 1 / 8, and 1 / 16 of the original input image, respectively. The EIOU regression loss function is used instead of CIOU to improve the model's ability to detect small targets.

[0025] The beneficial effects of this invention are:

[0026] This invention provides a target detection method for UAV aerial images based on spatial detail perception enhancement and cross-layer semantic fusion. First, a local feature information enhancement module is proposed to strengthen the local detail responses of multiple sets of feature maps at different scales, thereby improving feature representation capabilities. Second, a multi-scale feature aggregation network matching the model is proposed to achieve efficient fusion of feature maps at different scales, enhancing the global and local interaction of feature information. Finally, a spatial detail feature perception enhancement module is proposed to enhance the spatial details of the feature maps input to the detection head, improving the model's sensitivity to spatial location information. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings.

[0028] Figure 1 The overall flowchart of a target detection method for UAV aerial images based on spatial detail perception enhancement and cross-layer semantic fusion is provided by the present invention.

[0029] Figure 2 The diagram shows the structure of a target detection method for UAV aerial images based on spatial detail perception enhancement and cross-layer semantic fusion provided by this invention.

[0030] Figure 3 The multi-scale feature extraction network structure diagram provided by this invention;

[0031] Figure 4 A structural diagram of the local feature information enhancement module provided by this invention;

[0032] Figure 5 This is a structural diagram of the multi-scale feature aggregation network provided by the present invention;

[0033] Figure 6 This is a structural diagram of the lightweight bidirectional attention feature aggregation module provided by the present invention;

[0034] Figure 7 The network structure diagram of the spatial detail feature perception enhancement module provided by the present invention. Detailed Implementation

[0035] The technical solution will be further described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of the invention.

[0036] like Figure 1 As shown, this invention provides a target detection method for UAV aerial images based on spatial detail perception enhancement and cross-layer semantic fusion, comprising the following steps:

[0037] Step 1: Obtain the drone aerial image dataset and preprocess the acquired aerial images;

[0038] Step 2: Establish a target detection model for drone aerial images;

[0039] The input image is processed by a pre-defined multi-scale feature extraction network to extract features at multiple scales, thereby improving the model's global perception capability and obtaining four sets of feature maps A at different scales. i ;

[0040] For three sets of feature maps at different scales A iFeature enhancement is performed using a pre-defined local feature information enhancement module to strengthen the local feature information carried by feature maps at different scales, resulting in three sets of enhanced local information feature maps B at different scales. i ;

[0041] Feature map B is enhanced by using a pre-defined multi-scale feature aggregation network to enhance local information at three different scales. i Multi-scale feature aggregation is performed on the original feature map A4 at the smallest scale to obtain three sets of aggregated feature maps C at different scales. i ;

[0042] Three sets of aggregated feature maps at different scales C i After passing through the preset spatial detail perception enhancement module, three sets of enhanced spatial detail feature maps D are obtained. i ;

[0043] Three sets of enhanced spatial detail feature maps D i The input is fed into three preset detection heads for multi-target detection. The scales of the input feature maps corresponding to the detection heads are 1 / 4, 1 / 8 and 1 / 16 of the original input image, respectively, which correspond to the detection of small, medium and large targets.

[0044] Step 3: Train the model using the acquired drone aerial image dataset. After training and optimization, test the model using a test machine and obtain the test results.

[0045] Furthermore, in step 1, the VisDrone aerial image dataset is cleaned by deleting files with incorrect labels or corrupted images.

[0046] like Figure 2 As shown, the establishment of the UAV aerial image target detection model in step 2 is as follows: In the feature extraction stage, the input image is processed through a preset multi-scale feature extraction network to perform multi-scale feature extraction, thereby improving the model's global perception capability and obtaining four sets of feature maps A at different scales. i For three sets of feature maps at different scales, A i Feature enhancement is performed using a pre-defined local feature information enhancement module to strengthen the local feature information carried by feature maps at different scales, resulting in three sets of enhanced local information feature maps B at different scales. i Feature map B is enhanced by using a pre-defined multi-scale feature aggregation network to enhance local information at three different scales. i Multi-scale feature aggregation is performed on the original feature map A4 at the smallest scale to obtain three sets of aggregated feature maps C at different scales. i Three sets of aggregated feature maps at different scales C i After passing through the preset spatial detail perception enhancement module, three sets of enhanced spatial detail feature maps D are obtained. iThe three sets of enhanced spatial detail feature maps D i The input is fed into three preset detection heads for multi-target detection. The scales of the input feature maps corresponding to the detection heads are 1 / 4, 1 / 8 and 1 / 16 of the original input image, respectively, which correspond to the detection of small, medium and large targets.

[0047] like Figure 3 As shown, the multi-scale feature extraction network consists of a shallow feature extraction module, a hierarchical feature compression module, a high semantic awareness module, and a spatial self-attention enhancement module, aiming to achieve semantic detail extraction and information aggregation under different receptive fields. The input image first passes through two cascaded CBS modules (convolution + BN + SiLU activation) to extract initial edge and texture features and generate the first-stage feature map A1; then, local context modeling is performed through the C3K2 module, and spatial compression and semantic enhancement are combined with the SCDown module to extract intermediate features such as A2 and A3 step by step. This structure adopts a multi-level C3K2 module stacking design, which effectively deepens the network's receptive field without introducing too many parameters.

[0048] At the network's end, SPPF (Spatial Pyramid Pooling Fast Module) is used to perform multi-scale context fusion of deep features, further enhancing the network's ability to perceive targets at different scales. Following this, C2PSA (Dual-Branch Channel Spatial Attention Module) is introduced to achieve joint attention weighting across spatial and channel dimensions, thereby highlighting key target regions and generating the final highly semantically expressive feature map A4. This feature extraction network features progressively decreasing resolution and layer-by-layer semantic enhancement, effectively supporting subsequent detection heads to perform efficient and accurate multi-target detection in complex scenes.

[0049] like Figure 4 As shown, the local feature information enhancement module comprises five parallel branches. Branches 1 to 4 utilize convolutional combinations with different depths and dilation rates to capture multi-scale contextual information from local to global levels. Branch 1 uses shallow convolutions to preserve original spatial details; branches 2 to 4 expand the receptive field through progressively deeper convolutional structures and the introduction of dilated convolutions, thereby more effectively integrating rich contextual semantic information; branch 5 serves as a residual path to enhance the stability of feature fusion. This module first performs multiple 1×1 convolutions on the input feature map to reduce the number of channels and perform preliminary feature compression. Then, features are extracted through the five branches with different structures:

[0050] The first branch uses only 1×1 convolution;

[0051] The second branch is a 1×1 convolution followed by a standard 3×3 convolution, with a hole ratio d = 1;

[0052] The third branch consists of a 1×1 convolution followed by a standard 3×3 convolution, and finally a 3×3 dilated convolution with a dilation rate of d=3.

[0053] The fourth branch is a concatenation of 1×1 convolution followed by 1×3 convolution + 3×1 convolution, forming an approximate large convolution kernel structure, and finally a 3×3 dilated convolution with a dilation rate d = 5.

[0054] The fifth branch is the original input features followed by a 1×1 convolution.

[0055] The output features from the first to fourth branches are concatenated along the channel dimension, and then a 1×1 convolution is used to fuse the feature information. After fusion, a residual connection is introduced, adding the fifth branch to the fused feature map to achieve feature preservation and enhancement, outputting the final enhanced feature map. The implementation definition process of this module is as follows:

[0056] Branch 1 = Conv(A i )

[0057] Branch 2 = AConv(Conv(A) i ))

[0058] Branch 3 = AConv(Conv(Conv(A) i )))

[0059] Branch 4 = AConv(Conv(Conv(Conv(A)) i ))))

[0060] Branch 5 = Conv(A i )

[0061] B i = branch5 + Conv(Concat(Branch1 + branch2 + branch3 + branch4))

[0062] In the formula, A i B represents the multi-scale features extracted by the feature extraction network. i The output features are enhanced by the module. Conv represents Conv2d+BN+SiLU, Concat represents channel concatenation, and AConv represents dilated convolution. The dilation rate is {1, 3, 5} according to the branch depth to achieve a gradually expanding receptive field.

[0063] like Figure 5As shown, the multi-scale feature aggregation network takes four sets of features (B1, B2, B3, A4) at different scales output by the backbone network as input. This module adopts a dual-path structure of "bottom-up fusion + top-down enhancement," effectively integrating multi-scale features through a cascaded combination of the LBiFPN-ATT bidirectional attention feature fusion module, the C3K2 module, and CBS downsampling. Finally, it outputs three levels of aggregated features (C1, C2, C3) to adapt to the detection needs of targets at different scales. To better handle the four sets of input features at different scales, the network structure incorporates an additional upsampling layer and adapts and expands the channel dimensions. During aggregation, while maintaining the fine-grained feature information at the bottom layer, it reduces excessive semantic fusion of shallow features, enhances the semantic expressive power of deep features, and thus improves the model's ability to jointly perceive global context and local details.

[0064] like Figure 6 As shown, the bidirectional attention feature fusion module adopts a bottom-up and top-down bidirectional feature flow structure, combined with the lightweight spatial attention mechanism SimAM, to achieve efficient fusion and salient region enhancement among multi-scale features. This module receives input feature maps at multiple scales. First, it transmits high-level semantic information through a low-to-up path and fuses it with low-level semantic features. The fused features are further enhanced by the SimAM module to strengthen the target region response. Subsequently, the module uses a top-to-down path to transmit information and combines it with the enhanced features from the previous stage to complete a more context-aware fusion, again supplemented by the SimAM attention mechanism to improve spatial discriminability.

[0065] like Figure 7 As shown, the Spatial Detail-aware FeatureEnhancement Module is specifically designed to enhance the model's ability to perceive spatial details, and is particularly suitable for scenes with blurred boundaries of small targets, complex structures, or strong background interference. The module structure is as follows:

[0066] First, input feature C i The feature is initially processed using DSConv (Depthwise Separable Convolution) to preserve key spatial structure information while reducing the number of parameters and computational overhead. Subsequently, this feature is fed into three parallel feature enhancement branches to capture spatial details from multiple perspectives:

[0067] Branch1 uses an ACBS module (Atrous Convolution+BN+SiLU) with a hole ratio of 2 to expand the receptive field, enabling the network to acquire a wider range of contextual information while maintaining spatial resolution, which helps to perceive the semantic environment of small targets.

[0068] Branch 2 introduces the GroupConv (group convolution) structure, which enhances the model's ability to distinguish fine-grained spatial patterns by grouping and processing features from different channels, and extracts rich details from multiple subspaces;

[0069] Branch 3 introduces DeformConv (deformable convolution), which adaptively adjusts the sampling position of standard convolution, enabling the network to more flexibly perceive irregular boundaries and local deformable structures, thereby modeling the target edge more accurately.

[0070] The output features of the three branches are element-wise summed, fusing multiple types of spatial detail information. Subsequently, the lightweight attention mechanism SimAM is introduced to further enhance the network's focus on key regions (such as edges and texture abrupt changes), improving the accuracy of spatial response and suppressing invalid background interference. The overall structure integrates multi-scale receptive fields, diverse convolution methods, and an adaptive spatial attention mechanism, significantly enhancing the model's ability to express and understand the spatial details of the target while maintaining module efficiency, providing a more solid feature foundation for subsequent accurate detection. The implementation definition process of this module is as follows:

[0071] x = DSConv(C i )

[0072] Branch 1 = ACBS(Conv(x))

[0073] Branch 2 = GroupConv(Conv(x))

[0074] Branch 3 = DeformConv(Conv(x))

[0075] D i =SimAM(Branch1 + branch2 + branch3)

[0076] In the formula, C i D represents the input features. i The output features are represented by SimAM, an attention method, ACBS dilated convolution (dilation rate of 2), GroupConv group convolution, and DeformConv deformable convolution.

[0077] To verify the effectiveness of the above embodiments, the VisDrone dataset was used to evaluate the method provided in this invention. The average detection accuracy of the method for various types of targets was calculated and comprehensively compared with advanced methods for target detection in UAV aerial images, including advanced algorithms such as Super-YOLO, FFCA-YOLO, and YOLOv13.

[0078] The VisDrone dataset includes 6471 training images, 548 validation images, and 1610 test images collected by drones. The dataset contains ten object categories: pedestrians, humans, bicycles, cars, vans, trucks, tricycles, covered tricycles, buses, and motorcycles. Experimental results are shown in Table 1.

[0079] Table 1 Performance evaluation of different methods on the VisDrone test set.

[0080]

[0081] Experimental results show that the method proposed in this invention achieves the highest average detection accuracy among all comparison algorithms. Specifically, mAP 50 The proposed algorithm achieves a performance index (mAP) of 41.3, a 10.7% improvement over the baseline YOLOv11-m's 37.3; and a mAP of 25.0, an 11.6% improvement over the baseline's 22.4. In the car category, it achieves a detection accuracy of 80.1%, demonstrating the significant potential of the proposed algorithm in low-altitude traffic monitoring. Although the proposed method performs slightly worse than other algorithms in identifying covered tricycles (ATri), the overall detection results indicate that the proposed method has superior performance.

[0082] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A target detection method for UAV aerial images based on spatial detail perception enhancement and cross-layer semantic fusion, characterized in that, The method includes the following steps: Step 1: Data acquisition and preprocessing. Obtain the drone aerial image dataset and preprocess it. Step 2: Construct the object detection model; In the feature extraction stage, four sets of feature maps A at different scales were extracted from the high-level to the low-level using a pre-set multi-scale feature extraction network. i The feature maps are defined as follows: a bottom-level feature map A1, a middle-level feature map A2, a high-level feature map A3, and a deep-level feature map A4. The bottom-level feature map A1, the middle-level feature map A2, and the high-level feature map A3 are enhanced with local feature information through a preset local feature information enhancement module. The local feature information enhancement module includes a parallel multi-branch structure of standard convolutional path and multi-scale dilated convolutional path. It fuses local information under different receptive fields through residual connection and enhances the expressive ability of small target edge structure and context information, generating bottom-level feature map B1, middle-level feature map B2, and high-level feature map B3 with local feature enhancement, respectively. In the feature fusion stage, three sets of fused feature maps C are obtained through a pre-set multi-scale feature fusion network. i ; The three sets of locally enhanced feature maps B i Using a deep feature map A4 as input, this system innovatively constructs three bottom-up upsampling paths and two lateral feature transfer paths for multi-scale feature fusion. A pre-defined lightweight bidirectional attention feature aggregation module replaces the traditional upsampling and stitching structure to enhance the representational power of the fused features. A pre-defined spatial detail perception enhancement module enhances the model's ability to focus on spatial location and edge texture in the feature map, enabling more accurate perception of small targets. The lightweight bidirectional attention feature aggregation module consists of an attention mechanism and a constructed two-feature input bidirectional feature fusion network. The spatial detail perception enhancement module is a three-branch heterogeneous structure composed of depthwise separable convolution, dilated convolution, group convolution, deformable convolution, and an attention mechanism. After the fused feature map undergoes spatial detail perception enhancement through this module, three sets of spatially enhanced feature maps D are obtained. i ; The lightweight bidirectional attention feature aggregation module includes a multi-scale feature input branch, a cross-scale bidirectional feature aggregation path, and an embedded SimAM attention enhancement unit. The multi-scale feature input branch contains at least two parallel initial feature flow channels, and each input feature map completes preliminary feature extraction and transformation through sequentially connected feature processing nodes. The cross-scale bidirectional feature aggregation path constructs a hierarchical aggregation and information complementarity mechanism for cross-scale features through bidirectional connections between feature processing nodes, realizing bidirectional flow optimization of shallow detail features and deep semantic features. SimAM attention is deployed at the output end of the feature processing nodes and in the cross-scale feature transmission path. With its spatial detail perception capability, it adaptively weights and modulates the feature map, highlighting key spatial detail information of the target, suppressing redundant background features, driving the efficient fusion of multi-scale features in the bidirectional aggregation process, and enhancing the feature representation's ability to discriminate targets in UAV aerial photography scenarios. During the prediction phase, three sets of detector heads, configured at 1 / 4, 1 / 8, and 1 / 16, respectively receive three sets of spatial detail-enhanced feature maps D. i Construct a loss function to better address the characteristics of small targets being small in scale and densely packed. Step 3: Train the model using the preprocessed drone aerial image dataset, and test the best trained model on the test set to obtain the test results.

2. The target detection method for UAV aerial images based on spatial detail perception enhancement and cross-layer semantic fusion according to claim 1, characterized in that: The four sets of feature maps extracted by the multi-scale feature extraction network are 160×160, 80×80, 40×40 and 20×20 respectively.

3. The target detection method for UAV aerial images based on spatial detail perception enhancement and cross-layer semantic fusion according to claim 1, characterized in that: The local feature information enhancement module includes multiple parallel branches. Each branch uses standard convolution and dilated convolution with different dilation rates of 1, 3 and 5 for feature extraction. The output features of each branch are fused in the channel dimension and introduced into the original feature stream through residual connection to enhance the feature map's ability to perceive small target details and contextual information, thereby improving the detection accuracy of small targets in complex scenes.

4. The target detection method for UAV aerial images based on spatial detail perception enhancement and cross-layer semantic fusion according to claim 1, characterized in that: The spatial detail perception enhancement module includes three heterogeneous enhancement paths, based on depthwise separable convolution, ACBS module, grouped convolution, and deformable convolution, respectively, to model long-range dependencies, local structural features, and geometric deformation information in the image. The output features of the enhancement paths are fused and then introduced with the SimAM attention mechanism to achieve joint feature enhancement of the spatial and channel domains, thereby improving detection accuracy. The ACBS module is composed of dilated convolution and BN+SiLU.

5. The target detection method for UAV aerial images based on spatial detail perception enhancement and cross-layer semantic fusion according to claim 1, characterized in that: The feature maps received by the three sets of detection heads have resolutions of 160×160, 80×80, and 40×40, respectively; the loss function consists of two parts: classification loss and bounding box loss; the classification loss uses the binary cross-entropy loss function; the bounding box loss uses the Focal loss and EIOU loss functions.

Images