Unmanned aerial vehicle image small target detection method and device based on biaxial feature interaction

Abstract

According to the unmanned aerial vehicle image small target detection method and device based on biaxial feature interaction, four context awareness interaction stages (CAI Stage) are cascaded, and each stage comprises two context awareness interaction blocks (CAI Blocks). And each CAI Block realizes feature extraction through a biaxial parallel architecture, and performs feature interaction by adopting a key element fusion module (KEF Module) to finally form a detection network with a self-adaptive multi-scale perception capability. According to the CAI-DETR provided by the invention, through an innovative architecture design and a feature interaction mechanism, the constructed biaxial one-dimensional convolution architecture decouples feature learning in horizontal and vertical directions, and calculation redundancy is greatly reduced while the receptive field coverage capability is ensured. The lightweight design enables the network to be more focused on the extraction of the essential features of the target, and especially for the ubiquitous tiny target in the unmanned aerial vehicle image, the coupling relationship between the local texture and the overall contour of the tiny target is enhanced through an adaptive sampling strategy.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to a method, device and storage medium for detecting small targets in drone images based on dual-axis feature interaction. Background Art

[0002] Current drone image object detection technology primarily relies on convolutional neural networks (CNNs) and Transformer-based models. Traditional CNN methods (such as the YOLO series and Faster R-CNN) improve detection capabilities through multi-scale feature pyramids and attention mechanisms. However, they suffer from the following problems when detecting small objects: Insufficient fusion of local and global features: Existing models simply handle the interaction between local details and global background (such as direct addition or splicing), resulting in the dilution of key target features by background noise, making it difficult to balance detail preservation and context integration.

[0003] The contradiction between computational efficiency and accuracy: Mainstream methods use large-scale 2D convolution or complex attention mechanisms (such as Vision Transformer) to improve feature extraction capabilities, but with the expansion of parameter scale (such as ResNet-34 with 30.2M parameters), it is difficult to meet the real-time detection needs of drones.

[0004] Directional feature degradation: Existing single-axis convolution or fixed sampling strategies (such as 3×3 convolution) lead to the weakening of the correlation between horizontal and vertical features, which is prone to missed detection or false detection, especially under complex background interference.

[0005] Transformer-based detection frameworks (such as DETR and RT-DETR) enhance global modeling capabilities through the self-attention mechanism, but still have the following drawbacks: Sparsity of small object features: The Transformer's self-attention mechanism is insufficient for extracting low-resolution small object features, resulting in decreased detection accuracy (for example, DINO's mAP50 in the VisDrone2019 dataset is only 46.2%).

[0006] High hardware resource consumption: The computational complexity of large-scale self-attention increases quadratically (for example, the DINO model reaches 279 GFLOPS), making it difficult to deploy on drone platforms with limited computing power.

[0007] Existing improvements (such as LSKA-DETR and Drone-YOLO) attempt to mitigate these issues through dynamic convolution or lightweight design. However, core issues such as directional feature separation, local-global interaction efficiency, and small object background suppression remain unresolved. For example, LSKA-DETR uses a large-kernel attention mechanism, and while its parameters are limited to 20.8M, its mAP50 is only 46.7%, and it does not optimize for feature conflicts among densely packed small objects in drone imagery. Summary of the Invention

[0008] The present invention proposes a method, device and storage medium for detecting small targets in drone images based on dual-axis feature interaction, which can solve at least one of the technical problems in the background technology.

[0009] To achieve the above object, the present invention adopts the following technical solutions: A method for detecting small targets in UAV images based on dual-axis feature interaction is implemented by a computer device, S1. Build a dual-axis parallel feature extraction framework with drone image data as input; independently extract features through horizontal and vertical dual-axis branches: the horizontal branch uses 1×3 depthwise separable convolution DWConv to process the input image and generate a horizontal feature map; the vertical branch uses 3×1 depthwise separable convolution to process the input image and generate a vertical feature map; S2 performs a preliminary fusion of the initial features of the dual-axis branches; the horizontal feature map output by S1 is interacted with the vertical feature map through the key element fusion (KEF) module: first, 1×1 convolution is performed on the features of the two branches to adjust the channel dimension, and then nonlinear fusion is achieved through the ReLU activation function and element-by-element product. Finally, the channels are spliced and the fused features are output through the convolution layer, with the dimension consistent with the input; S3. Input the fused features into the context-aware interaction (CAI) module. Based on the adaptive dual-path context-aware extraction (ADCE) mechanism, it is decomposed into two paths: local feature extraction and global feature extraction. The local path uses non-dilated convolution with a dilation rate of 1 (1×3 / 3×1) to capture small target details. The global path uses multi-branch 1D deformable dilated convolution with a dilation rate of 2 / 3 / 3 and kernel size of 1×5 / 1×7 / 1×11 to expand the receptive field and extract background context information. S4: Enhance the local and global features output by S3; Use AG-Attention anchoring to guide the element-by-element product attention to dynamically fuse dual-path features: After summing the local and global features, perform maximum pooling and average pooling respectively to generate a spatial descriptor, and generate a query vector Q through 1×1 convolution; Local features are convolved to generate keys K and values V; Use element-by-element product to calculate the attention weight, normalize it with Softmax, and weight it with value V to output the enhanced features; S5: Perform secondary fusion on the enhanced dual-axis features. The AG-Attention outputs of the horizontal and vertical axes output by S4 are input into the key element fusion (KEF) module again. Cross-axis feature interaction is achieved through element-by-element multiplication and channel concatenation, and the fused high-order features are output. S6. Build a multi-stage context-aware interaction network. Repeat the S1-S5 process four times, corresponding to four CAI stages. Each stage includes two dual-axis feature extraction and fusion processes, including two KEF modules. The input and output of each stage are superimposed through residual connections to gradually deepen the feature representation. S7. Output the detection results; input the final fused features into the RT-DETR detection head, generate the target bounding box prediction through the Transformer encoder-decoder structure, and combine the bipartite graph matching loss to achieve end-to-end training.

[0010] Furthermore, step S2 includes setting up a context-aware interaction network CAI-DETR. The backbone network of the context-aware interaction network CAI-DETR is composed of four independent context-aware interaction stages CAI Stages. Each CAI stage contains two context-aware interaction blocks CAI Blocks, and each context-aware interaction block CAI Block is integrated with two CAI modules and two KEF modules. Specifically, the CAI modules on each axial branch integrate the contextual information around the target object through the ADCE mechanism, and use the AG-Attention mechanism to enhance target-related features while suppressing irrelevant background interference. The KEF module effectively fuses the features extracted by different axial branches through a parallel architecture and element-by-element multiplication operations.

[0011] Furthermore, in the context-aware interaction network CAI-DETR architecture, the context-aware interaction block serves as the core feature extraction unit, which implements feature modeling in complex scenarios through a dual-axis parallel processing mechanism; This module adopts a two-stage iterative processing process to input the feature map X l Feature extraction and fusion are performed on two independent branches, the horizontal axis and the vertical axis, and finally a feature representation with global context perception is formed through residual connection; In the first iteration, the input feature X l Directional feature extraction is performed through dual-axis depthwise separable convolution DWConv; The horizontal branch uses a 1×3 DWConv kernel to capture horizontal features and generate a horizontal feature map F ; The vertical branch uses a 3×1 DWConv kernel to extract vertical features and obtain a vertical feature map ; Subsequently, the key element fusion module KEF Module performs nonlinear interaction on the dual-axis features, establishes cross-axis feature associations through the element-by-element product operation of the channel dimension, and combines the fusion results with the original input X l Perform residual connection to generate intermediate feature X l' ,This design effectively alleviates the gradient vanishing problem of deep networks; In the second iteration, the intermediate feature X l Entering the enhancement processing stage; each axial branch performs multi-scale feature extraction through the context-aware interaction module (CAI Module), which includes the adaptive dual-path context extraction mechanism ADCE Mechanism and the anchor-guided element-by-element attention mechanism AG-Attention; The ADCE mechanism adopts a local-global dual-path architecture: the local path uses 1×3 / 3×1 standard convolution kernels to extract target details; the global path uses multi-scale dilated convolutions of 1×5, 1×7, and 1×11 to build a large receptive field and obtain contextual information around the target; The features of the two paths are dynamically fused through AG-Attention. This mechanism uses spatial pooling features to generate a query vector, performs element-by-element product interaction with the key-value pairs generated by local features, and generates attention weights through softmax normalization, ultimately achieving targeted enhancement of local features based on global context. The dual-axis features processed by the CAI module are input into the KEF module again for feature fusion; The KEF module in this stage adopts a parallel dual-branch structure to perform horizontal feature AF. and vertical feature AF After performing 1×1 convolution and ReLU activation respectively, element-by-element multiplication operations across branches are performed, and the final fusion features are generated through channel splicing and convolution dimensionality reduction; both iterative processes retain the residual connection structure to ensure the integrity of the underlying features.

[0012] Furthermore, S3 proposes an adaptive dual-path context-aware extraction mechanism, ADCE Mechanism, through local-global feature decoupling and dynamic sampling strategy. This mechanism adopts a dual-path parallel architecture, where the local path focuses on capturing fine-grained target features, while the global path focuses on building large-scale context associations. The two achieve complementary enhancement through a parameter-adaptive feature fusion strategy. The local feature extraction path uses 1D deformable convolution technology, deploying 1×3 and 3×1 asymmetric convolution kernels on the horizontal and vertical axes, respectively. By introducing a learnable offset parameter, the spatial distribution of the convolution kernel sampling points can be dynamically adjusted based on the input features, allowing the feature extractor to accurately adapt to the morphological diversity of targets under the drone's perspective. The global feature extraction path innovatively adopts a multi-branch dilated convolution structure. To address the spatial correlation between targets and backgrounds in drone images, three sets of asymmetric dilated convolution kernels are designed: 1×5 / d=2, 1×7 / d=3, and 1×11 / d=3 in the horizontal direction. Through a layered dilation strategy, the feature receptive field expands exponentially, with a maximum coverage area of 33×33 pixels. The global path adopts the "non-center point sampling" initialization strategy, and the sampling point set is defined as R ={(x_i,0)|x_i∈Z}, avoiding target feature contamination caused by overlapping center points; The dual-path features are dynamically fused through the anchor-guided attention mechanism AG-Attention. The high-resolution feature map output by the local path retains the target geometric details, while the contextual semantic information provided by the global path is modulated by the spatial weight matrix.

[0013] Furthermore, in the attention mechanism AG-Attention, an adaptive feature interaction mechanism is proposed. This mechanism aims to enhance the representation ability of local features by integrating multi-scale background and contextual information. It uses global contextual information to dynamically adjust the importance weights of different regions in the local feature map, thereby improving the model's attention to key target areas. Specifically, the input of AG-Attention is the horizontal direction of the local features extracted by the ADCE mechanism. and global features First, the local and global features are added element by element, and the spatial descriptor is extracted through the maximum pooling MaxPool and average pooling AvgPool operations in the channel dimension: =MaxPool( + ) =AvgPool( + ) Subsequently, the two pooling results are added together and passed through a 1x1 convolution to generate the query vector Q, while the key vector K and value vector V are directly obtained from the local features through a 1x1 convolution: Q = Conv1x1( + ) K = Conv1x1( ) V = Conv1x1( ) AG-Attention uses element-wise product instead of matrix multiplication to calculate the attention score. This method captures the correlation between local features and global context through point-by-point interaction, significantly reducing the computational complexity: = Softmax(Q ⊙ K) ⊙ V Where ⊙ represents the element-wise product operation, and Softmax normalizes the attention weights along the channel dimension; Furthermore, S5 specifically includes: In the key element fusion module (KEF), a feature interaction fusion strategy based on parallel structure and element-by-element product is proposed to solve the problem of feature spatial position degradation caused by the limited receptive field when extracting features in the horizontal and vertical directions using single-axis one-dimensional convolution. The KEF module achieves efficient information complementation while maintaining feature independence by introducing a parallel feature interaction mechanism and an element-by-element nonlinear fusion strategy. In the specific implementation, given the horizontal feature map F And vertical feature map F First, the channel dimension is adjusted through 1x1 convolution to adapt to subsequent operations, and the ReLU activation function is used to introduce nonlinear transformation capabilities: = ReLU(Conv1×1( ))⊙Conv1×1( ) = ReLU(Conv1×1( ))⊙Conv1×1( ) Where ⊙ represents the element-wise product operation; Then, the interactive horizontal features With vertical features Splicing along the channel dimension and further fusing multi-scale information through the convolution layer: =Conv(Concat( , )).

[0014] In another aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0015] On the other hand, the present invention further discloses a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0016] It can be seen from the above technical solution that the method for detecting small targets in drone images based on dual-axis feature interaction of the present invention includes inputting drone images and performing standardized preprocessing, uniformly scaling the original images to 640×640 resolution, performing normalization processing to eliminate light interference, and constructing batch input tensors; extracting directional features through a dual-axis parallel convolution architecture, using 1×n deformable convolution for horizontal feature sampling on the horizontal axis, and using n×1 deformable convolution for vertical feature sampling on the vertical axis, respectively generating local feature maps and global context feature maps with dynamic offsets; performing context-aware feature interaction, spatially aligning the horizontal and vertical axis features, and then calculating cross-axis correlations through an anchor-guided element-by-element product attention mechanism. The weights are normalized using Softmax2D and then weighted fused with the original features; multi-level nonlinear feature fusion is implemented, and after 1×1 convolution dimension reduction of the dual-axis features, the cross-axis feature correlation is enhanced through bidirectional product interaction, and the fusion feature map is generated after splicing along the channel dimension and depth-separable convolution; a backbone network is constructed, and multi-scale features are extracted step by step through four context-aware interaction stages. Each stage contains two dual-axis feature extraction and fusion operations, and residual connections are used to retain the original feature information; the end-to-end detection head is used to output the results, and the final feature map is input into the RT-DETR detection head to generate predicted box coordinates and category probabilities. The model is optimized based on the bipartite matching loss function, and the small target detection results in the drone image are output.

[0017] The CAI-DETR proposed in this paper demonstrates significant advantages in the field of drone image target detection through its innovative architecture design and feature interaction mechanism. Different from the parameter-intensive characteristics of traditional two-dimensional convolutional networks, the dual-axis one-dimensional convolutional architecture constructed in this work decouples feature learning in the horizontal and vertical directions, significantly reducing computational redundancy while ensuring receptive field coverage. This lightweight design allows the network to focus more on extracting the essential features of the target, especially for the small targets that are prevalent in drone images. The coupling relationship between their local texture and overall contour is enhanced through an adaptive sampling strategy.

[0018] The core ADCE mechanism breaks through the rigid sampling pattern of traditional convolution kernels and dynamically adjusts the feature extraction area through a deformable offset. In the local path, the convolution kernel can autonomously focus on the core target area (such as pedestrian head outlines and vehicle edge features), effectively suppressing complex background interference. The global path captures the semantic relationship between the target and the surrounding environment through an asymmetric sampling strategy, forming a multi-granular context perception capability. This dynamic balancing mechanism demonstrates unique advantages in densely populated small target scenes. Experimental results show that it can significantly reduce false detection rates and improve target positioning accuracy.

[0019] At the feature enhancement level, AG-Attention innovatively uses anchor-guided element-by-element product attention to inject global semantic information into local feature representations. Unlike traditional attention mechanisms that rely on dot product operations and have high computational complexity, this method enhances the response intensity of the target area through a spatial modulation strategy, demonstrating greater robustness in extreme scenarios such as low light and target occlusion. At the same time, the KEF module achieves deep cross-axis feature fusion through dual-path parallel product interaction, overcoming the semantic conflict problem in traditional serial fusion. This interaction mechanism not only preserves the independence of multi-directional features, but also selectively enhances target-related features through nonlinear gating, effectively improving target discrimination in dense scenes.

[0020] Visual comparisons and quantitative experiments demonstrate that this method can adapt to the dynamic perspective changes and complex background interference of drone imagery while maintaining its lightweight characteristics. Its dual-path collaborative design philosophy inherits the advantages of convolutional networks' sensitivity to local features while incorporating global contextual reasoning capabilities, providing a new technical path for achieving high-precision real-time detection on embedded drone platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A schematic diagram of a context-aware interaction network according to an embodiment of the present invention; Figure 2 A schematic diagram of a context-aware interaction block according to an embodiment of the present invention; Figure 3 Schematic diagram of the adaptive context-aware attention mechanism according to an embodiment of the present invention; Figure 4 Schematic diagram of an anchor-guided attention mechanism implemented in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0023] like Figure 1 As shown in the figure, the method for detecting small targets in drone images based on dual-axis feature interaction described in this embodiment proposes a context-aware dual-axis feature extraction framework (CAI-DETR) for drone images, which solves the problem of small target detection in complex backgrounds through innovative modular design. Figure 1As shown in the figure, the system consists of four cascaded context-aware interaction stages (CAI stages), each of which contains two context-aware interaction blocks (CAI blocks). Each CAI block extracts features through a dual-axis parallel architecture and uses a key element fusion module (KEF module) for feature interaction, ultimately forming a detection network with adaptive multi-scale perception capabilities.

[0024] Specifically, it includes the following steps: S1. Build a dual-axis parallel feature extraction framework with drone image data as input. Independently extract features through horizontal and vertical dual-axis branches: the horizontal branch processes the input image using 1×3 depthwise separable convolution (DWConv) to generate a horizontal feature map; the vertical branch processes the input image using 3×1 depthwise separable convolution to generate a vertical feature map. S2. Perform preliminary fusion of the initial features of the dual-axis branches. The horizontal feature map output by S1 is interacted with the vertical feature map through the key element fusion (KEF) module: first, 1×1 convolution is performed on the features of the two branches to adjust the channel dimension, and then nonlinear fusion is achieved through the ReLU activation function and element-by-element product. Finally, the channels are concatenated and the fused features are output through the convolution layer, with the dimension consistent with the input. S3. The fused features are input into the context-aware interaction (CAI) module. Based on the adaptive dual-path context-aware extraction (ADCE) mechanism, this is decomposed into two paths: local feature extraction and global feature extraction. The local path uses non-dilated convolutions with a dilation rate of 1 (1×3 / 3×1) to capture small object details. The global path uses multi-branch 1D deformable dilated convolutions with dilation rates of 2 / 3 / 3 and kernel sizes of 1×5 / 1×7 / 1×11 to expand the receptive field and extract background context information. S4: Enhance the local and global features output by S3. AG-Attention anchoring guides element-wise product attention to dynamically fuse dual-path features: sum the local and global features, perform maximum pooling and average pooling respectively to generate a spatial descriptor, and generate the query vector Q through 1×1 convolution; local features are subjected to 1×1 convolution to generate keys K and values V; the attention weights are calculated using element-wise product, normalized by Softmax, and weighted with the value V to output the enhanced features; S5: Perform secondary fusion on the enhanced dual-axis features. The AG-Attention outputs of the horizontal and vertical axes output by S4 are input into the key element fusion KEF module again, and cross-axis feature interaction is achieved through element-by-element product and channel splicing, and the fused high-order features are output; S6. Build a multi-stage context-aware interaction network. Repeat S1-S5 four times, corresponding to four CAI stages. Each stage includes two dual-axis feature extraction and fusion processes, including two KEF modules. The input and output of each stage are superimposed through residual connections, gradually deepening the feature representation. S7. Output detection results. The final fused features are fed into the RT-DETR detection head, which generates target bounding box predictions through the Transformer encoder-decoder structure. End-to-end training is achieved by combining bipartite graph matching loss.

[0025] To address the challenges brought by the diversity of target scales and background complexity in UAV target detection, this paper innovatively proposes a context-aware interaction network (CAI-DETR). Figure 1 As shown in Figure 1, the CAI-DETR backbone network consists of four independent context-aware interaction stages (CAI stages). Each CAI stage contains two context-aware interaction blocks (CAIBlocks), and each CAI block integrates two CAI modules and two KEF modules. Specifically, the CAI modules on each axial branch incorporate contextual information around the target object through the ADCE mechanism and employ the AG-Attention mechanism to enhance target-related features while suppressing irrelevant background interference. The KEF module effectively fuses features extracted from different axial branches through a parallel architecture and element-wise product operations.

[0026] The core architecture includes the following key modules: 1. Context-Aware Interaction Blocks (CAI Blocks) In the CAI-DETR architecture, the Context-Aware Interaction Block serves as the core feature extraction unit, which implements feature modeling in complex scenarios through a dual-axis parallel processing mechanism. Figure 2 As shown, this module adopts a two-stage iterative processing flow to input feature map X l Feature extraction and fusion are performed on two independent branches, the horizontal axis and the vertical axis, respectively, and finally a feature representation with global context perception is formed through residual connection.

[0027] In the first iteration, the input feature X l Directional features are extracted through dual-axis depthwise separable convolution (DWConv). The horizontal branch uses a 1×3 DWConv kernel to capture horizontal directional features and generate a horizontal feature map. ; The vertical branch uses a 3×1 DWConv kernel to extract vertical features and obtain a vertical feature map This asymmetric convolution kernel design effectively enhances the ability to extract direction-sensitive features while maintaining parameter efficiency. Subsequently, the key element fusion module (KEF Module) performs nonlinear interaction on the dual-axis features, establishes cross-axis feature associations through element-by-element product operations in the channel dimension, and combines the fusion results with the original input X l Perform residual connection to generate intermediate feature X l' ,This design effectively alleviates the gradient vanishing problem of deep networks.

[0028] In the second iteration, the intermediate feature X l Entering the enhancement processing stage. Each axial branch performs multi-scale feature extraction through the context-aware interaction module (CAI Module), which includes an adaptive dual-path context extraction mechanism (ADCE Mechanism) and an anchor-guided element-by-element attention mechanism (AG-Attention). The ADCE mechanism adopts a local-global dual-path architecture: the local path uses 1×3 / 3×1 standard convolution kernels to extract target detail features; the global path uses multi-scale dilated convolution (1×5, 1×7, 1×11) to construct a large receptive field to obtain contextual information around the target. The features of the two paths are dynamically fused through AG-Attention. This mechanism uses spatial pooling features to generate a query vector, which is element-wise multiplied with the key-value pairs generated by the local features. Attention weights are generated through softmax normalization, ultimately achieving targeted enhancement of local features by the global context.

[0029] The dual-axis features processed by the CAI module are input into the KEF module again for feature fusion. In this stage, the KEF module adopts a parallel dual-branch structure to process the horizontal features. and vertical features After performing 1×1 convolution and ReLU activation, a cross-branch element-wise product operation is performed, and the final fused features are generated through channel concatenation and convolution dimensionality reduction. Both iterative processes retain the residual connection structure to ensure the integrity of the underlying features. This two-stage processing architecture significantly enhances the model's feature representation of small objects in drone imagery while maintaining parameter efficiency through a progressively refined feature interaction strategy. In particular, directional feature decomposition effectively alleviates the feature confusion problem of traditional two-dimensional convolution in complex backgrounds.

[0030] Adaptive Dual-Path Context-Aware Extraction Mechanism (ADCE Mechanism) In the task of detecting small targets in complex drone images, traditional convolutional neural networks face significant challenges due to their fixed sampling patterns and limited receptive fields. The Adaptive Dual-Path Context-Aware Extraction Mechanism (ADCE Mechanism) proposed by CAI-DETR effectively overcomes these limitations through the innovative design of local-global feature decoupling and dynamic sampling strategies. Figure 3 As shown in the figure, the mechanism adopts a dual-path parallel architecture, in which the local path focuses on capturing fine-grained features of the target, while the global path focuses on building large-scale contextual associations. The two achieve complementary enhancement through a parameter-adaptive feature fusion strategy. The local feature extraction path adopts 1D deformable convolution technology, deploying 1×3 and 3×1 asymmetric convolution kernels for the horizontal and vertical axes respectively. By introducing a learnable offset parameter, the convolution kernel sampling points can dynamically adjust the spatial distribution according to the input features, so that the feature extractor can accurately adapt to the morphological diversity of the target from the perspective of the drone. For example, when processing in the horizontal direction, the initial sampling point set Defined as {(x_i,0)|x_i∈Z}, the offset Δp optimized by backpropagation i This allows the convolution kernel to flexibly scale along the x-axis, effectively capturing the continuous features of narrow, elongated objects (such as vehicle queues). Compared to traditional 2D deformable convolutions, this design reduces the number of offset parameters by approximately 60%, achieving feature-sensitive optimization along a single axis while maintaining a lightweight model.

[0031] The global feature extraction path innovatively adopts a multi-branch dilated convolution structure. Three sets of asymmetric dilated convolution kernels (1×5 / d=2, 1×7 / d=3, and 1×11 / d=3 in the horizontal direction) are designed to address the spatial correlation characteristics of the target and background in drone images. Through the layered dilation strategy, the feature receptive field expands exponentially, with a maximum coverage area of 33×33 pixels. Of particular note, the global path uses a "non-center point sampling" initialization strategy, defining the sampling point set as ={(x_i,0)|x_i∈Z}, avoiding target feature contamination caused by center point overlap. This design enables the global path to focus on background context modeling. For example, in dense crowd detection, the interference pattern of individual targets and group shadows can be distinguished through long-range dependencies.

[0032] The dual-path features are dynamically fused through an anchor-guided attention mechanism (AG-Attention). The high-resolution feature maps output by the local path preserve the target geometric details, while the contextual semantic information provided by the global path is modulated by a spatial weight matrix.

[0033] 2. Context-Aware Interaction Block (CAI Block): In the Anchor-Guided Entrywise ProductAttention (AG-Attention) module, this paper proposes an adaptive feature interaction mechanism that aims to enhance the representational capabilities of local features by integrating multi-scale background and contextual information. The core concept of this module is to use global context information to dynamically adjust the importance weights of different regions in the local feature map, thereby improving the model's focus on key target areas. Traditional methods that simply add or concatenate local and global features can lead to redundancy or information conflict. AG-Attention effectively solves this problem through its innovative attention calculation method.

[0034] Specifically, the input of AG-Attention is the local features extracted by the ADCE mechanism (such as the horizontal ) and global features (such as First, the local and global features are added element by element, and the spatial descriptor is extracted through the maximum pooling (MaxPool) and average pooling (AvgPool) operations of the channel dimension: =MaxPool( + ) =AvgPool( + ) Subsequently, the two pooling results are added together and passed through a 1x1 convolution to generate the query vector Q, while the key vector K and value vector V are directly obtained from the local features through a 1x1 convolution: Q = Conv 1x1 ( + ) K = Conv 1x1 ( ) V = Conv 1x1 ( ) Unlike traditional attention mechanisms, AG-Attention uses element-wise product (EEP) instead of matrix multiplication to calculate attention scores. This method captures the correlation between local features and global context through point-by-point interaction, significantly reducing computational complexity: = Softmax(Q ⊙ K) ⊙ V Here, ⊙ represents an element-wise product operation, and Softmax normalizes the attention weights along the channel dimension. This design enables the model to adaptively enhance regions of local features that are highly relevant to the global context while suppressing irrelevant background noise. Ultimately, the product of the attention weights and the local feature map further enhances the response strength of the target area. This mechanism can effectively improve feature sensitivity to low-resolution targets, especially in small object detection tasks.

[0035] The advantage of AG-Attention lies in its balance between computational efficiency and feature fusion. Compared to the traditional Transformer attention mechanism, the element-by-element product strategy reduces the number of parameters while maintaining flexibility in feature selection, enabling the model to more accurately locate targets in complex drone imagery.

[0036] In the Key Element Fusion Module (KEF module), a feature interaction fusion strategy based on parallel structure and element-by-element product is proposed to address the problem of feature spatial position degradation caused by the limited receptive field when extracting features in the horizontal and vertical directions using single-axis one-dimensional convolution. The core goal of this module is to compensate for the shortcomings of the receptive field in a single direction by efficiently fusing the features of the horizontal and vertical branches, while strengthening the semantic relevance of cross-directional features. Although traditional cascade operations can retain multi-directional feature information, they may introduce redundancy and cause a surge in channel dimensions, while simple feature addition (Sum) makes it difficult to capture complex cross-directional feature relationships. The KEF module achieves efficient information complementarity while maintaining feature independence by introducing a parallel feature interaction mechanism and an element-by-element nonlinear fusion strategy.

[0037] In the specific implementation, given the horizontal feature map and vertical feature maps First, the channel dimension is adjusted through 1x1 convolution to adapt to subsequent operations, and the ReLU activation function is used to introduce nonlinear transformation capabilities: = ReLU(Conv 1×1 ( ))⊙Conv 1×1 ( ) = ReLU(Conv 1×1 ( ))⊙Conv 1×1 ( ) Where ⊙ represents an element-by-element product operation. This step uses asymmetric cross-directional product interactions to mutually enhance high-response regions in the horizontal features with corresponding regions in the vertical features, while suppressing low-response regions (such as background noise), thus forming a feature enhancement mechanism with directionality.

[0038] Then, the interactive horizontal features With vertical features Splicing along the channel dimension and further fusing multi-scale information through the convolution layer: =Conv(Concat( , )) This design replaces traditional linear operations with nonlinear interactions using element-by-element multiplication, reducing the number of parameters while enhancing the collaborative expression of cross-directional features. For example, when a spatial location exhibits high activation in both horizontal and vertical features, its product result will significantly increase the response value of that location, thereby highlighting the key target area. Conversely, if the response of a particular directional feature is weak, the product operation will automatically reduce the fusion weight of that location, effectively suppressing background interference.

[0039] Compared to cascade operations, the KEF module achieves adaptive feature selection and enhancement through product interaction, avoiding the computational burden caused by channel expansion. Through its lightweight structural design, this module significantly improves the model's detection robustness for small, direction-sensitive targets, especially in scenes such as building edges and dense crowds, showing stronger feature discrimination.

[0040] The implementation method takes the VisDrone2019 dataset detection as an example. The specific implementation process is as follows: Input preprocessing: scale the drone image to 640×640 resolution and perform normalization Feature extraction: Stages 1-4 are processed sequentially, and each CAI Stage contains 2 CAI Blocks The horizontal branch uses 1×3 deformable convolution, and the vertical branch uses 3×1 deformable convolution. The global path in ADCE Mechanism sets three sets of asymmetric dilated convolutions Feature fusion: Primary fusion uses the element-by-element product strategy of the KEF module The final output is generated by the RT-DETR detection head to generate a prediction box Training optimization: Use Adam optimizer with an initial learning rate of 0.0001 The loss function uses the standard bipartite matching loss of the DETR series Alternatives Convolution implementation alternative: ADCE Mechanism can use 2D deformable convolution, but it will increase the number of parameters by 31.1%. The global path can be replaced by the Atrous Spatial Pyramid Pooling (ASPP) structure Attention mechanism alternative: AG-Attention can use the Transformer architecture instead, but the computational complexity increases by 14.5%. You can try out the combination of channel attention (SE Block) and spatial attention Fusion module replacement: The KEF module can adopt the traditional solution of channel splicing + 3×3 convolution You can try to add a gating mechanism to control the feature fusion weight VI. Experimental To comprehensively evaluate the effectiveness of each module of CAI-DETR, we designed three sets of ablation experiments. The first set of experiments evaluated the independent impact of each module by gradually adding modules to the baseline model LSKA-DETR. Compared with the baseline, CAI-DETR improved mAP(0.5) and mAP(0.5:0.95) by 2.9% and 2.0%, respectively, with only a slight increase in the number of parameters and GFLOPS. The second and third sets of experiments respectively evaluated the impact of different implementations of the ADCE mechanism and AG-Attention on model performance.

[0041] The second set of experimental results demonstrates that implementing the ADCE mechanism using one-dimensional deformable convolution achieves the best balance between performance and computational cost. The third set of experiments shows that compared to the summation operation, the proposed AG-Attention implementation improves mAP (0.5) by 2.0% and mAP (0.5:0.95) by 0.9%.

[0042] Table 1 shows the results of the first set of ablation experiments conducted on the VisDrone2019 dataset. It can be seen that when either the ADCE mechanism or the KEF module is added alone, both mAP (0.5) and mAP (0.5:0.95) are improved, with minimal increase in the number of parameters. When the ADCE mechanism and the KEF module are used together, the performance improvement is significantly better than adding either module alone. The combination of the MSFE module and the KEF module achieves the best results, with mAP (0.5) and mAP (0.5:0.95) increased by 0.7% and 0.5%, respectively, compared to using the MSFE module alone.

[0043] Table 1 Ablation experiment results of different modules on VisDrone2019-val

[0044] The CAI-DETR proposed in this paper demonstrates significant advantages in the field of drone image target detection through its innovative architecture design and feature interaction mechanism. Different from the parameter-intensive characteristics of traditional two-dimensional convolutional networks, the dual-axis one-dimensional convolutional architecture constructed in this work decouples feature learning in the horizontal and vertical directions, significantly reducing computational redundancy while ensuring receptive field coverage. This lightweight design allows the network to focus more on extracting the essential features of the target, especially for the small targets that are prevalent in drone images. The coupling relationship between their local texture and overall contour is enhanced through an adaptive sampling strategy.

[0045] The core ADCE mechanism breaks through the rigid sampling pattern of traditional convolution kernels and dynamically adjusts the feature extraction area through a deformable offset. In the local path, the convolution kernel can autonomously focus on the core target area (such as pedestrian head outlines and vehicle edge features), effectively suppressing complex background interference. The global path captures the semantic relationship between the target and the surrounding environment through an asymmetric sampling strategy, forming a multi-granular context perception capability. This dynamic balancing mechanism demonstrates unique advantages in densely populated small target scenes. Experimental results show that it can significantly reduce false detection rates and improve target positioning accuracy.

[0046] At the feature enhancement level, AG-Attention innovatively uses anchor-guided element-by-element product attention to inject global semantic information into local feature representations. Unlike traditional attention mechanisms that rely on dot product operations and have high computational complexity, this method enhances the response intensity of the target area through a spatial modulation strategy, demonstrating greater robustness in extreme scenarios such as low light and target occlusion. At the same time, the KEF module achieves deep cross-axis feature fusion through dual-path parallel product interaction, overcoming the semantic conflict problem in traditional serial fusion. This interaction mechanism not only preserves the independence of multi-directional features, but also selectively enhances target-related features through nonlinear gating, effectively improving target discrimination in dense scenes.

[0047] Visual comparisons and quantitative experiments demonstrate that this method can adapt to the dynamic perspective changes and complex background interference of drone imagery while maintaining its lightweight characteristics. Its dual-path collaborative design philosophy inherits the advantages of convolutional networks' sensitivity to local features while incorporating global contextual reasoning capabilities, providing a new technical path for achieving high-precision real-time detection on embedded drone platforms.

[0048] In another aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0049] On the other hand, the present invention further discloses a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0050] In another embodiment provided in the present application, a computer program product comprising instructions is also provided, which, when executed on a computer, enables the computer to execute any of the methods for detecting small targets in drone images based on dual-axis feature interaction in the above-mentioned embodiments.

[0051] It is understandable that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts of the above methods.

[0052] In the above embodiments, all or part of the embodiments can be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, hard disk, tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0053] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0054] Each embodiment in this specification is described in a related manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiment is generally similar to the method embodiment, so the description is relatively simple. For related parts, refer to the description of the method embodiment.

[0055] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A small target detection method for UAV images based on dual-axis feature interaction, characterized in that: The following steps are included: S1. Build a dual-axis parallel feature extraction framework with drone image data as input; independently extract features through horizontal and vertical dual-axis branches: the horizontal branch uses 1×3 depthwise separable convolution DWConv to process the input image and generate a horizontal feature map; the vertical branch uses 3×1 depthwise separable convolution to process the input image and generate a vertical feature map; S2 performs a preliminary fusion of the initial features of the dual-axis branches; the horizontal feature map output by S1 is interacted with the vertical feature map through the key element fusion (KEF) module: first, 1×1 convolution is performed on the features of the two branches to adjust the channel dimension, and then nonlinear fusion is achieved through the ReLU activation function and element-by-element product. Finally, the channels are spliced and the fused features are output through the convolution layer, with the dimension consistent with the input; S3. Input the fused features into the context-aware interaction (CAI) module. Based on the adaptive dual-path context-aware extraction (ADCE) mechanism, it is decomposed into two paths: local feature extraction and global feature extraction. The local path uses non-dilated convolution with a dilation rate of 1 (1×3 / 3×1) to capture small target details. The global path uses multi-branch 1D deformable dilated convolution with a dilation rate of 2 / 3 / 3 and kernel size of 1×5 / 1×7 / 1×11 to expand the receptive field and extract background context information. S4, perform attention enhancement on the local features and global features output by S3; The AG-Attention anchor guides the element-wise product attention to dynamically fuse dual-path features: After summing the local and global features, they are respectively subjected to maximum pooling and average pooling to generate a spatial descriptor, and the query vector Q is generated through a 1×1 convolution; the local features are subjected to a 1×1 convolution to generate the key K and value V; the attention weight is calculated using the element-wise product, normalized by Softmax, and weighted with the value V to output the enhanced feature; S5: Perform secondary fusion on the enhanced dual-axis features. The AG-Attention outputs of the horizontal and vertical axes output by S4 are input into the key element fusion (KEF) module again. Cross-axis feature interaction is achieved through element-by-element multiplication and channel concatenation, and the fused high-order features are output. S6. Build a multi-stage context-aware interaction network. Repeat the S1-S5 process four times, corresponding to four CAI stages. Each stage includes two dual-axis feature extraction and fusion processes, including two KEF modules. The input and output of each stage are superimposed through residual connections to gradually deepen the feature representation. S7. Output the detection results; input the final fused features into the RT-DETR detection head, generate the target bounding box prediction through the Transformer encoder-decoder structure, and combine the bipartite graph matching loss to achieve end-to-end training.

2. The method for detecting small targets in drone images based on dual-axis feature interaction according to claim 1, characterized in that: Step S2 includes setting up a context-aware interaction network CAI-DETR. The backbone network of the context-aware interaction network CAI-DETR consists of four independent context-aware interaction stages CAI Stages. Each CAI stage contains two context-aware interaction blocks CAI Blocks, and each context-aware interaction block CAI Block integrates two CAI modules and two KEF modules. Specifically, the CAI modules on each axial branch integrate the contextual information around the target object through the ADCE mechanism, and use the AG-Attention mechanism to enhance target-related features while suppressing irrelevant background interference. The KEF module effectively fuses the features extracted by different axial branches through a parallel architecture and element-by-element multiplication operations.

3. The method for detecting small targets in drone images based on dual-axis feature interaction according to claim 2, characterized in that: In the context-aware interaction network CAI-DETR architecture, the context-aware interaction block serves as the core feature extraction unit, achieving feature modeling in complex scenarios through a dual-axis parallel processing mechanism; This module adopts a two-stage iterative processing process to input the feature map X l Feature extraction and fusion are performed on two independent branches, the horizontal axis and the vertical axis, and finally a feature representation with global context perception is formed through residual connection; In the first iteration, the input feature X l Directional feature extraction is performed through dual-axis depthwise separable convolution DWConv; The horizontal branch uses a 1×3 DWConv kernel to capture horizontal features and generate a horizontal feature map F ; The vertical branch uses a 3×1 DWConv kernel to extract vertical features and obtain a vertical feature map ; Subsequently, the key element fusion module KEF Module performs nonlinear interaction on the dual-axis features, establishes cross-axis feature associations through the element-by-element product operation of the channel dimension, and combines the fusion results with the original input X l Perform residual connection to generate intermediate feature X l' ,This design effectively alleviates the gradient vanishing problem of deep networks; In the second iteration, the intermediate feature X l Entering the enhancement processing stage; each axial branch performs multi-scale feature extraction through the context-aware interaction module (CAI Module), which includes the adaptive dual-path context extraction mechanism ADCE Mechanism and the anchor-guided element-by-element attention mechanism AG-Attention; The ADCE mechanism adopts a local-global dual-path architecture: the local path uses 1×3 / 3×1 standard convolution kernels to extract target details; the global path uses multi-scale dilated convolutions of 1×5, 1×7, and 1×11 to build a large receptive field and obtain contextual information around the target; The features of the two paths are dynamically fused through AG-Attention. This mechanism uses spatial pooling features to generate a query vector, performs element-by-element product interaction with the key-value pairs generated by local features, and generates attention weights through softmax normalization, ultimately achieving targeted enhancement of local features based on global context. The dual-axis features processed by the CAI module are input into the KEF module again for feature fusion; The KEF module in this stage adopts a parallel dual-branch structure to perform horizontal feature AF. and vertical feature AF After performing 1×1 convolution and ReLU activation respectively, the cross-branch element-wise product operation is performed, and the final fusion feature is generated through channel splicing and convolution dimensionality reduction; Both iterative processes retain the residual connection structure to ensure the complete transmission of underlying features.

4. The method for detecting small targets in drone images based on dual-axis feature interaction according to claim 3 is characterized by: S3 proposes an adaptive dual-path context-aware extraction mechanism, ADCE Mechanism, through local-global feature decoupling and dynamic sampling strategy. This mechanism adopts a dual-path parallel architecture, where the local path focuses on capturing fine-grained target features, while the global path focuses on building large-scale context associations. The two achieve complementary enhancement through a parameter-adaptive feature fusion strategy. The local feature extraction path uses 1D deformable convolution technology, deploying 1×3 and 3×1 asymmetric convolution kernels on the horizontal and vertical axes, respectively. By introducing a learnable offset parameter, the spatial distribution of the convolution kernel sampling points can be dynamically adjusted based on the input features, allowing the feature extractor to accurately adapt to the morphological diversity of targets under the drone's perspective. The global feature extraction path innovatively adopts a multi-branch dilated convolution structure. To address the spatial correlation between targets and backgrounds in drone images, three sets of asymmetric dilated convolution kernels are designed: 1×5 / d=2, 1×7 / d=3, and 1×11 / d=3 in the horizontal direction. Through a layered dilation strategy, the feature receptive field expands exponentially, with a maximum coverage area of 33×33 pixels. The global path adopts the "non-center point sampling" initialization strategy, and the sampling point set is defined as R ={(x_i,0)|x_i∈Z}, avoiding target feature contamination caused by overlapping center points; The dual-path features are dynamically fused through the anchor-guided attention mechanism AG-Attention. The high-resolution feature map output by the local path retains the target geometric details, while the contextual semantic information provided by the global path is modulated by the spatial weight matrix.

5. The method for detecting small targets in drone images based on dual-axis feature interaction according to claim 4 is characterized in that: In the attention mechanism AG-Attention, an adaptive feature interaction mechanism is proposed. It aims to enhance the representation ability of local features by integrating multi-scale background and contextual information. It uses global context information to dynamically adjust the importance weights of different regions in the local feature map, thereby improving the model's attention to key target areas. Specifically, the input of AG-Attention is the horizontal direction of the local features extracted by the ADCE mechanism. and global features First, the local and global features are added element by element, and the spatial descriptor is extracted through the maximum pooling MaxPool and average pooling AvgPool operations in the channel dimension: =MaxPool( + ) =AvgPool( + ) Subsequently, the two pooling results are added together and passed through a 1x1 convolution to generate the query vector Q, while the key vector K and value vector V are directly obtained from the local features through a 1x1 convolution: Q = Conv 1x1 ( + ) K = Conv 1x1 ( ) V = Conv 1x1 ( ) AG-Attention uses element-wise product instead of matrix multiplication to calculate the attention score. This method captures the correlation between local features and global context through point-by-point interaction, significantly reducing the computational complexity: = Softmax(Q ⊙ K) ⊙ V Where ⊙ represents the element-wise product operation, and Softmax normalizes the attention weights along the channel dimension.

6. The method for detecting small targets in drone images based on dual-axis feature interaction according to claim 5, characterized in that: S5 specifically includes: In the key element fusion module (KEF), a feature interaction fusion strategy based on parallel structure and element-by-element product is proposed to solve the problem of feature spatial position degradation caused by the limited receptive field when extracting features in the horizontal and vertical directions using single-axis one-dimensional convolution. The KEF module achieves efficient information complementation while maintaining feature independence by introducing a parallel feature interaction mechanism and an element-by-element nonlinear fusion strategy. In the specific implementation, given the horizontal feature map F and vertical feature maps First, the channel dimension is adjusted through 1x1 convolution to adapt to subsequent operations, and the ReLU activation function is used to introduce nonlinear transformation capabilities: = ReLU(Conv 1×1 ( ))⊙Conv 1×1 ( ) = ReLU(Conv 1×1 ( ))⊙Conv 1×1 ( ) Where ⊙ represents the element-wise product operation; Then, the interactive horizontal feature KF With vertical feature KF Splicing along the channel dimension and further fusing multi-scale information through the convolution layer: =Conv(Concat( , ))。 7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 6.

Images