Unmanned aerial vehicle small target detection method based on improved YOLO26s
By improving the backbone and neck networks of the YOLO26s model, introducing windmill-shaped wavelet transform convolution, A2C2f modules, and DySample dynamic upsampling, and combining them with a context anchor residual attention feature fusion module, the problem of feature information loss in small target detection on UAV platforms is solved, and efficient small target detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- QUANZHOU INST OF EQUIP MFG
- Filing Date
- 2026-03-18
- Publication Date
- 2026-06-02
AI Technical Summary
Existing YOLO26s models struggle to accurately identify and locate small targets in drone scenarios, primarily due to oversampling of the backbone network, lack of dedicated mechanisms for preserving small target features, and the absence of attention mechanisms, leading to the loss of feature information. Furthermore, lightweight convolutional schemes are difficult to balance detection performance on drone platforms with limited computing resources.
In the YOLO26s model, a windmill-shaped wavelet transform convolution module, an A2C2f module, and DySample dynamic upsampling are introduced. Combined with a context anchor residual attention feature fusion module, the feature extraction and fusion paths of the backbone and neck networks are optimized to enhance the detection capability of small targets.
It significantly improves the accuracy and robustness of small target detection, reduces computational complexity, adapts to the computational resource limitations of UAV platforms, and achieves efficient small target detection.
Smart Images
Figure CN121904640B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of small target detection, and specifically to a method for detecting small targets on unmanned aerial vehicles (UAVs) based on an improvement of YOLO26s. Background Technology
[0002] YOLO26s represents a significant upgrade over its predecessors, notably by removing the DFL module to reduce computational overhead and enhance hardware compatibility, implementing end-to-end NMS-free inference to simplify post-processing and improve deployment robustness, and introducing the ProgLoss loss function and STAL label allocation strategy to improve training stability and small target recognition accuracy, thus significantly improving small target detection performance. This model combines the advantages of a streamlined structure and fast inference speed, meeting the real-time requirements of various scenarios while ensuring high detection accuracy. However, when applied to drone scenarios, it still faces significant challenges due to the combined effects of complex background interference, extremely low target pixel ratio, and insufficient image resolution. As a general-purpose single-stage model, its small target detection performance is severely limited, making it difficult to accurately identify and locate small targets.
[0003] Existing models, including YOLO26s, generally lack a dedicated mechanism for preserving features of small targets. On the one hand, the traditional feature pyramid flow design is prone to oversampling of the backbone network, and the feature fusion link between the backbone network and the neck is inefficient in transferring and fusing fine-grained features of small targets, resulting in the loss of feature information. On the other hand, mainstream target detection architectures and the native design of YOLO26s do not deploy attention mechanisms specifically adapted to small targets, causing the model to ignore the extraction of salient features of small targets in the shallow network stage, further weakening the ability to detect small targets.
[0004] Drones are mobile embedded platforms, and their computing units are subject to strict constraints in terms of power consumption and computing power. Existing lightweight downsampling convolutional solutions, such as ODConv and GhostConv, even if they can reduce the number of model parameters or computational load by more than 20%, are generally accompanied by a significant decline in various detection accuracy indicators, making it difficult to balance lightweight design and detection performance, and thus unsuitable for the actual application needs of drone scenarios. Summary of the Invention
[0005] The purpose of this invention is to provide a UAV small target detection method based on YOLO26s that improves detection performance and adapts to UAV application scenarios.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A method for small target detection in UAVs based on an improved YOLO26s model is proposed. The method constructs a YOLO26s-SOD model for small target detection in UAVs. The YOLO26s-SOD model is an improved YOLO26s model. The YOLO26s-SOD model includes a backbone network for feature extraction of input images, a neck network for feature extraction and fusion of feature maps, and a head structure for detection and classification of the fused feature maps output by the neck network. Skip connections are set between the backbone network and the neck network.
[0008] Replace the first and second C3k2 modules of the original backbone network with context anchor residual attention feature fusion modules, replace the third C3k2 module with an A2C2f module, replace the convolution module between the context anchor residual attention feature fusion module and the corresponding A2C2f module with a windmill wavelet transform convolution module, delete the modules after the windmill wavelet transform convolution module and connect them to the first convolution module;
[0009] The first Upsample module in the original neck network is replaced with the first connection module, the context anchor residual attention feature fusion module, and the DySample module connected in sequence. The second Upsample module is replaced with the DySample module. The C3k2 module in the neck network is replaced with the context anchor residual attention feature fusion module. After the second C3k2 module, the DySample module, the second connection module, and the context anchor residual attention feature fusion module are added in sequence.
[0010] The images captured by the drone are input into the YOLO26s-SOD model, and the detection results are output.
[0011] Preferably, the skip connection includes a second convolutional module, a third convolutional module, a fourth convolutional module, a fifth convolutional module, and a sixth convolutional module;
[0012] The second convolutional module is used to perform a convolution operation on the features output by the first convolutional module of the backbone network and then feed them to the second connection module.
[0013] The third convolutional module is used to perform a convolution operation on the features output by the first context anchor residual attention feature fusion module of the backbone network and then feed them to the second connection module of the neck network.
[0014] The fourth convolutional module is used to perform a convolution operation on the features output by the first context anchor residual attention feature fusion module of the backbone network and then feed them to the first connection module of the neck network.
[0015] The fifth convolutional module is used to perform a convolution operation on the features output by the second context anchor residual attention feature fusion module of the backbone network and then feed them to the first connection module.
[0016] The sixth convolutional module is used to perform a convolution operation on the features output by the second context anchor residual attention feature fusion module of the neck network and then feed them to the last connection module of the neck network.
[0017] Preferably, the windmill-shaped wavelet transform convolution module converts the feature map... Input the residual connection branch and the channel expansion branch respectively;
[0018] This residual connection branch will connect the feature map Input a 1×1 convolutional module, perform channel adjustment and downsampling, and output residual connection features;
[0019] The channel expansion branch pairs with the feature map. Adjust the number of channels to obtain the feature map. A multi-level wavelet transform module is used to process the feature map. Processing: The first-level wavelet transform module transforms the feature map Decomposed into Sub-band Sub-band Sub-band and The four sub-bands are concatenated and then subjected to joint feature modulation via a channel-grouped 1×1 convolutional module. The output is also divided into four groups of enhanced sub-bands. , , and The four sets of enhanced subband input inverse wavelet transform modules are reconstructed back to the original spatial resolution to generate multi-scale fused feature maps. ;
[0020] Expanding the branch also includes the feature map Input convolution module for feature extraction to obtain convolutional features ;
[0021] Multi-scale fusion feature maps With convolutional features Add and fuse to generate wavelet-enhanced feature maps ;
[0022] Four sets of depthwise convolutions were used to enhance the wavelet feature map. Multi-directional spatial feature extraction is performed, and the output features are added element-wise to complete the initial fusion of directional features. The fusion results are then sequentially processed through batch normalization and activation functions to generate directionally enhanced spatial features. ;
[0023] The wavelet-enhanced feature map Spatial features Element-wise addition is performed, followed by 1×1 convolution for channel integration, and the enhanced features are output.
[0024] The enhancement feature is added element-wise to the residual connection feature to output the first feature map.
[0025] Preferably, the context anchor residual attention feature fusion module performs channel expansion on the input features to obtain expanded features. These expanded features are used for branch selection according to configuration parameters, with the specific rules as follows:
[0026] When the configuration parameter attn=True, multiple SequentiaIList modules are used to process this extended feature. The SequentiaIList module includes the PSABlock-CARA module and the Bottleneck-CARA module connected in sequence.
[0027] When the configuration parameter c3k=True, the C3k-CARA module is used to process this extended feature;
[0028] When both configuration parameters c3k and attn are False, the extended feature is processed using Bottleneck-CARA.
[0029] The extended feature is then concatenated with the features output from each branch, and the concatenated feature map is output via a convolution module.
[0030] By adopting the aforementioned design scheme, the beneficial effects of this invention are as follows: This application improves the YOLO26s model by introducing a windmill-shaped wavelet transform convolution module in the downsampling stage of the backbone network. It integrates wavelet transform and multi-directional windmill-shaped convolution to simultaneously extract frequency domain details and spatial context information. By integrating these four sets of convolution results, a windmill-shaped receptive field distribution is formed as a whole. All convolutions adopt a grouped convolution method, which effectively controls the computational complexity while maintaining directional characteristics.
[0031] The A2C2f module is introduced into the neck network. The core of the A2C2f module consists of multiple ABlock units. Each ABlock integrates an attention mechanism and a multilayer perceptron, forming an efficient dual-path feature processing architecture. This design significantly reduces the number of model parameters and computational complexity while enhancing the representation ability of multi-scale features, especially improving the accuracy of capturing key features of small targets. Removing the end downsampling layer effectively avoids the loss of small target details due to excessive reduction in spatial resolution, and provides high-resolution backbone features that retain more complete spatial structure information for subsequent neck fusion, thereby improving the overall effectiveness of the feature pyramid and the small target detection performance.
[0032] Introducing DySample dynamic upsampling into the neck network significantly reduces the computational burden of the model because DySample avoids complex dynamic convolution operations, making it very suitable for deployment on edge computing devices such as drones.
[0033] Optimize the multi-scale feature fusion path to enhance the semantic expression of shallow high-resolution features;
[0034] By embedding a context anchor residual attention feature fusion module into the backbone and neck networks, the dual-branch context anchor attention based on the iRMB residual structure and CAA evolution is integrated into the C3k2 architecture to form a configurable hierarchical attention mechanism, which significantly improves the spatial focusing ability of small targets. During feature extraction, this module can effectively establish spatial correlations between distant pixels, significantly expand the receptive field of the model, and thus enhance its ability to model long-distance spatial dependencies in wide-area scenes from the perspective of UAVs. Attached Figure Description
[0035] Figure 1 This is an architectural diagram of the YOLO26s-SOD model of the present invention;
[0036] Figure 2 This is a structural diagram of the A2C2f module of the present invention;
[0037] Figure 3 This is a structural diagram of the windmill-shaped wavelet transform convolution module of the present invention;
[0038] Figure 4 This is a structural diagram of the DySample module of the present invention;
[0039] Figure 5 This is a structural diagram of the context anchor residual attention module of the present invention;
[0040] Figure 6 This is a structural diagram of the context anchor residual attention feature fusion module of the present invention;
[0041] Figure 7 This is a visualization comparison of the detection effects of the present invention in different scenarios. Detailed Implementation
[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0043] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.
[0044] A method for small target detection in UAVs based on improved YOLO26s is proposed, which constructs a YOLO26s-SOD model specifically optimized for small target detection tasks in UAV scenarios. Figure 1 As shown, this YOLO26s-SOD model is an improved YOLO26s model.
[0045] The YOLO26s-SOD model includes a backbone network for feature extraction from the input image, a neck network for feature extraction and fusion of the feature map, and a head structure for detection and classification of the fused feature map output by the neck network. Skip connections are set between the backbone network and the neck network.
[0046] The first and second C3k2 modules of the original backbone network are replaced with context anchor residual attention feature fusion modules (C3k2CARA), the third C3k2 module is replaced with an A2C2f module, the convolution module between the context anchor residual attention feature fusion module and the corresponding A2C2f module is replaced with a windmill wavelet transform convolution module, and the modules after the windmill wavelet transform convolution module are deleted and connected to the first convolution module; in this embodiment, the first convolution module is a 1×1 convolution module.
[0047] The first Upsample module in the original neck network is replaced with a first connecting module, a context anchor residual attention feature fusion module, and a DySample module connected in sequence. The second Upsample module is replaced with a DySample module. The C3k2 module in the neck network is replaced with a context anchor residual attention feature fusion module. After the second C3k2 module, a DySample module, a second connecting module, and a context anchor residual attention feature fusion module are added in sequence. The image captured by the UAV is input into the YOLO26s-SOD model, and the detection results are output.
[0048] The skip connection includes a second convolutional module, a third convolutional module, a fourth convolutional module, a fifth convolutional module, and a sixth convolutional module;
[0049] The second convolutional module is used to perform a convolution operation on the features output by the first convolutional module of the backbone network and then feed them to the second connection module.
[0050] The third convolutional module is used to perform a convolution operation on the features output by the first context anchor residual attention feature fusion module of the backbone network and then feed them to the second connection module of the neck network.
[0051] The fourth convolutional module is used to perform a convolution operation on the features output by the first context anchor residual attention feature fusion module of the backbone network and then feed them to the first connection module of the neck network.
[0052] The fifth convolutional module is used to perform a convolution operation on the features output by the second context anchor residual attention feature fusion module of the backbone network and then feed them to the first connection module.
[0053] The sixth convolutional module is used to perform a convolution operation on the features output by the second context anchor residual attention feature fusion module of the neck network and then feed them to the last connection module of the neck network.
[0054] In this embodiment, the second and third convolutional modules are 1×1 convolutional modules, and the fourth, fifth, and sixth convolutional modules are 3×3 convolutional modules.
[0055] In this embodiment, as Figure 3As shown, the Pinwheel-Shaped WaveletTransform Convolution (PSWTConv) module integrates the advantages of Pinwheel-Shaped Convolution (PSConv) and Wavelet Transform Convolution (WTConv) to construct a highly efficient feature extraction unit specifically designed for small object detection. In terms of architecture, PSWTConv adopts a dual-branch residual structure. The specific processing flow of this Pinwheel-Shaped Wavelet Transform Convolution module is as follows:
[0056] The windmill-shaped wavelet transform convolution module converts the feature map Input the residual connection branch and the channel expansion branch respectively;
[0057] This residual connection branch will connect the feature map Input a 1×1 convolutional module, perform channel adjustment and downsampling, and output residual connection features;
[0058] In this embodiment, the residual connection branch dynamically selects a processing strategy based on the matching of the number of input and output channels C1 and C2 and the downsampling step size s: when C1 ≠ C2 or s > 1, a 1×1 convolution is used for channel adjustment and spatial downsampling; otherwise, the identity connection is maintained. The channel expansion branch also performs channel alignment: if C1 ≠ C2, a 1×1 convolution is used to adjust the channel dimension; if C1 = C2, the identity mapping is maintained.
[0059] The channel expansion branch pairs with the feature map. Adjust the number of channels to the target number C2 and obtain the feature map. A multi-level wavelet transform module is used to process the feature map. Processing: The first-level wavelet transform module transforms the feature map Decomposed into preserving global structure Sub-band, capturing horizontal edges Sub-band, capturing vertical edges Sub-bands and those that retain fine textures The four subbands are concatenated into a tensor of (4C2, H / 2, W / 2), and then subjected to joint feature modulation via a channel-grouped 1×1 convolutional module. The output is also divided into four groups of enhanced subbands. , , and This approach achieves cross-band information exchange and enhancement while maintaining channel independence. Low-frequency subband convolution focuses on preserving the global structure, while high-frequency subband convolution enhances edge details to meet the needs of small target detection.
[0060] The four sets of enhanced subband input inverse wavelet transform (IWT) modules are reconstructed back to the original spatial resolution to generate multi-scale fused feature maps. ;
[0061] Expanding the branch also includes the feature map Input convolution module for feature extraction to obtain convolutional features ;
[0062] Multi-scale fusion feature maps With convolutional features The features are added together and fused to generate wavelet-enhanced feature maps with rich frequency domain information. Wavelet-enhanced feature maps This will be retained directly as important contextual information in subsequent fusion stages. This feature map It retains the original spatial structure while incorporating multi-band details, and the size and number of channels remain unchanged throughout the process. It can be directly input into subsequent modules and is compatible with the residual connection and multi-scale fusion requirements of the YOLO architecture.
[0063] Four sets of depthwise convolutions (DWConv) were used to enhance the wavelet feature map. Multi-directional spatial feature extraction is performed, and the output features are added element-wise to complete the initial fusion of directional features. The fusion results are then sequentially activated by batch normalization and the SiLU activation function to generate directionally enhanced spatial features. ;
[0064] In this embodiment, the four groups of depthwise convolutions include two groups of horizontal convolutions (kernel size 1×k) and two groups of vertical convolutions (kernel size k×1). Each of the horizontal and vertical convolutions is configured with one base version and one cheaper version convolution branch, as detailed below:
[0065] The convolutional layer employs a built-in symmetrical Same Padding design. The horizontal basic branch has padding set to (0, k / / 2), padding only on the left and right sides of the horizontal dimension (k / / 2 pixels), with no padding in the vertical dimension. The vertical basic branch has padding set to (k / / 2, 0), padding only on the top and bottom sides of the vertical dimension (k / / 2 pixels), with no padding in the horizontal dimension. This ensures that the spatial dimensions of the feature map after convolution are consistent with those of the wavelet-enhanced feature map. completely consistent;
[0066] First, enhance the wavelet feature map Perform directional symmetric padding, i.e., padding the horizontal offset branch with k / 2 pixels to the left and right, and padding the vertical offset branch with k / 2 pixels to the top and bottom. After convolution, crop back the wavelet-enhanced feature map. The original dimensions are used to extract offset features outside the spatial location of the basic branches.
[0067] The wavelet-enhanced feature map Spatial features Element-wise addition is performed to achieve complementary fusion of frequency domain enhancement features and spatial orientation features. Then, 1×1 convolution is used for channel integration to enhance the feature representation and output the enhanced features. If the module downsampling stride s is greater than 1, average pooling is performed after channel integration to complete spatial downsampling.
[0068] The enhancement feature is added element-wise to the residual connection feature to output a first feature map with residual connection characteristics.
[0069] In summary, the windmill-shaped wavelet transform convolution module comprises four groups of directional convolutions, each employing a symmetrical padding strategy. Two groups are standard horizontal (1×k kernel) and vertical (k×1 kernel) convolutions (k=3 here), directly applied to the input feature map; the other two groups are applied in the same direction, but the feature map is first offset and padded before the convolution is performed. By fusing the results of these four convolutions, a windmill-like receptive field distribution is formed. All convolutions use a grouped convolution approach, effectively controlling computational complexity while preserving directional characteristics.
[0070] The core advantages of this design are: expanding the receptive field through windmill-shaped convolution to enhance the ability to capture contextual information of small targets; extracting multi-scale frequency domain features through wavelet transform to improve the richness of feature representation; and achieving efficient information flow and gradient propagation through a bi-branch residual structure. Applying the windmill-shaped wavelet transform convolution module to small target detection tasks can significantly improve the detection accuracy and robustness of the model in complex backgrounds, providing an effective technical path to solve problems such as easy loss of small target features and strong background interference.
[0071] An innovative Multi-Scale Hierarchical Corrected Feature Pyramid Network (MHCFPN) is proposed. Its core is to significantly improve multi-scale target detection performance through backbone network module replacement, downsampling layer simplification, and neck network fusion link reconstruction, with a particular emphasis on enhancing small target detection capabilities. The specific design is as follows:
[0072] In this embodiment, the following method is used for backbone network optimization: Figure 2The A2C2f module shown replaces the C2PSA and SPPF modules in the prototype network, and removes the last downsampling layer at the end of the backbone layer. The core of the A2C2f module consists of multiple ABlock units, each integrating an attention mechanism and a multilayer perceptron, forming an efficient dual-path feature processing architecture. This design significantly reduces the number of model parameters and computational complexity while enhancing the representation ability of multi-scale features, particularly improving the accuracy of capturing key features of small targets. Removing the end downsampling layer effectively avoids the loss of small target details due to excessive reduction in spatial resolution, providing high-resolution backbone features that retain more complete spatial structural information for subsequent neck fusion, thereby improving the overall effectiveness of the feature pyramid and the performance of small target detection.
[0073] In the neck fusion link design, multiple cross-level fusion links were added, along with an additional dynamic upsampling layer, DySample, forming a differentiated fusion architecture with three levels of upsampling and hierarchical downsampling outputs. The new fusion links break the traditional unidirectional fusion mode, enabling cross-scale interaction of features from different layers of the backbone network, downsampling from 2x to 16x. By accurately fusing shallow fine-grained features, mid-level structural features, and deep contextual features, the complementarity of features at each scale is enhanced, especially providing sufficient deep semantic information for small object detection. The additionally introduced DySample dynamic upsampling layer replaces conventional interpolation upsampling, adaptively preserving feature edges and details during scale transformation.
[0074] The core logic of this design is as follows: After refining the mid- and shallow features through the first and second upsampling, a third upsampling is performed to increase the feature scale to twice that of the ultra-shallow layer, directly outputting the corresponding detection head to enhance small target detection. Simultaneously, starting from the output features of the second upsampling, one and two downsampling operations are performed sequentially to complete feature scale correction and secondary enhancement, outputting feature maps adapted for mid- and deep target detection respectively, ultimately forming differentiated outputs from the three detection heads. This design achieves a closed-loop feature interaction of top-down refinement, bottom-up enhancement, and hierarchical output adaptation. It fully mines the features of shallow small targets through three-level upsampling, while ensuring the representational ability of mid- and deep target features through hierarchical downsampling, effectively solving the detection accuracy problems caused by weakened shallow features, scale mismatch, and insufficient detection head adaptability in traditional feature pyramids.
[0075] Overall, MHCFPN significantly improves the multi-scale feature integration and adaptation capabilities of the feature pyramid through a collaborative design that includes module replacement, downsampling layer simplification, addition of fusion links, and optimization of the upsampling output architecture. This network ensures both the contextual modeling capability of deep features and enhances the discriminative power of shallow, small target features, providing an efficient feature fusion solution for multi-scale target detection in complex scenarios.
[0076] In this embodiment, in the MHCFPN structure, DySample achieves upsampling by dynamically generating sampling positions instead of learning convolutional kernel weights, significantly simplifying the computation process. For example... Figure 4 As shown, the feature of size C×H×W The DySample module is input with a sample set S of size 2g×sH×sW. The sampling function grid_sample of the DySample module samples the features. Resample the sample set S to obtain the target upsampled features. During the sampling point generation process, input features First, an offset O is generated through linear and pixel rearrangement operations. This offset O is added to the reference coordinates of the original sampling function to form the sampling set S. Then, the sampling set S and the input features... Resampling is achieved through a sampling function. This module has a simple structure and is computationally efficient, making it particularly suitable for detecting densely packed small targets in UAV images. Because DySample avoids complex dynamic convolution operations, it significantly reduces the computational burden on the model, making it ideal for deployment on edge computing devices such as UAVs.
[0077] In this embodiment, the Context Anchor Residual Attention Feature Fusion Module (C3k2CARA) constructs a multi-layered variable feature enhancement network by fusing the C3k2 architecture with the Context Anchor Residual Attention mechanism. Based on the C3k2 structure in the YOLO detection framework, this module innovatively integrates the CARA attention mechanism into its core processing unit, forming three configurable enhancement paths to achieve adaptive feature extraction for targets at different scales.
[0078] like Figure 5 As shown, the Context Anchor Residual Attention (CARA) module integrates the core ideas of the Inverted Residual Block with the Context Anchor Attention (CAA) mechanism to construct a highly efficient feature enhancement unit specifically designed for intensive prediction tasks on mobile devices such as drones. This module effectively establishes spatial relationships between distant pixels during feature extraction, significantly expanding the model's receptive field and thus enhancing its ability to model long-distance spatial dependencies in wide-area scenes from a drone's top-down perspective.
[0079] This context anchor residual attention module inherits the efficient feature transformation capability of the inverted residual module and draws on the attention concept from the Transformer structure, making it particularly suitable for real-time intensive prediction tasks on mobile devices such as drones. The CARA module can perceive the information distribution of the entire input space during feature extraction, effectively expanding the model's receptive field and enhancing its ability to capture features of multi-scale targets from the drone's perspective. Specifically, the CARA module employs a hybrid network structure design, using deep convolution (DWConv) to efficiently capture local spatial features, while utilizing a CAA-based improved dual-branch context anchor attention module (DCAA) mechanism to model global spatial dependencies. The dual-branch context anchor attention module achieves long-distance modeling capabilities similar to traditional self-attention with lower computational cost through an innovative grouped strip convolution design. This collaborative design allows the model to simultaneously focus on local details and global structure, significantly enhancing its feature sensitivity to small targets, enabling it to accurately focus on low-pixel-ratio targets of interest in complex drone aerial backgrounds.
[0080] The core of the CARA module lies in its dual-path attention architecture. The CARA module first performs batch normalization on the input features, then extracts spatial reference anchor information through global average pooling, and finally captures long-distance spatial dependencies using horizontal and vertical bar-based DWConv convolutions. The horizontal direction uses a 1-row × multi-column convolution kernel, while the vertical direction uses a multi-row × 1-column convolution kernel. An asymmetric padding strategy is used to maintain the feature map size, and grouped convolutions achieve efficient feature extraction. The responses from the two directions are fused and integrated by a 1×1 convolution, then processed by a sigmoid function to generate a spatial attention weight map. This weight map is element-wise multiplied with the original feature map to achieve spatial recalibration of the features. Subsequently, a 3×3 DWConv captures local detail information, followed by a squeezing excitation mechanism to adaptively recalibrate the channel dimensions, enhancing the response of important channels. Next, the features are regularized by a Dropout layer to prevent overfitting, and then a 1×1 convolution adjusts the number of output channels. Finally, the transformed features are fused with the original input features through residual connections, thereby achieving effective information transfer and gradient flow. This dual-path design not only improves the module's adaptability to different tasks, but also enhances the model's expressive power while ensuring computational efficiency through reasonable structural design.
[0081] like Figure 6As shown, the main structure of this context anchor residual attention feature fusion module inherits the channel segmentation and feature fusion design of C3k2. This module uses a 1x1 convolution module to expand the input features through channels, obtaining expanded features. The key innovation of the C3k2CARA module lies in its three-layer variable processing architecture. This variable processing architecture dynamically configures the internal structure of the module through the Boolean parameter attn and c3k, achieving feature enhancement of different complexities. The expanded features select branches according to the configuration parameters, with the specific rules as follows:
[0082] When the configuration parameter attn=True, multiple SequentiaIList modules are used to process the extended feature. These SequentiaIList modules include the PSABlock-CARA and Bottleneck-CARA modules connected sequentially. Bottleneck-CARA embeds CARA attention into the standard bottleneck module, first performing feature transformation through two consecutive convolutions, and then applying CARA attention for spatial context modeling. PSABlock-CARA combines CARA attention with two MLP layers to form an enhancement processing unit, further strengthening the feature representation capability. This approach is suitable for deep networks requiring strong context modeling.
[0083] When the configuration parameter c3k=True, the C3k-CARA module is used to process the extended feature. The C3k-CARA module uses Bottleneck-CARA as the basic component to build a cascaded processing sequence containing multiple Bottleneck-CARA, forming a multi-scale feature interaction network with attention enhancement, balancing feature richness and computational efficiency.
[0084] When both configuration parameters c3k and attn are False, the extended feature is processed using Bottleneck-CARA. This path only uses the basic CARA attention mechanism and is suitable for shallow networks or resource-constrained environments where computational efficiency is critical.
[0085] The extended feature is then concatenated with the features output from each branch along the channel dimension, and the concatenated feature map is output through a 1x1 convolution module.
[0086] In its implementation, the C3k2CARA module achieves adaptive architecture selection through a dynamic parameter configuration mechanism. Based on the network depth and task requirements, the system can pre-configure the parameters for each layer during training: an attention enhancement path is enabled in the deep feature extraction stage to strengthen long-distance dependency modeling; a C3k ensemble path is used in the mid-level feature processing stage to balance performance and efficiency; and a lightweight basic path is used in the shallow feature extraction stage to minimize computational overhead. This hierarchical attention mechanism design allows the model to adaptively adjust the receptive field range and feature interaction intensity at different network depths.
[0087] This application validates the effectiveness of the proposed YOLO26s-SOD model through systematic ablation experiments and multi-dimensional comparisons. As shown in Table 1, based on the YOLO26s baseline (mAP@0.5 = 39.9%, number of parameters 9.47M, GFLOPs = 20.5), introducing MHCFPN (Model A) alone brings a 5.9 percentage point increase in mAP@0.5 to 45.8%, while significantly compressing the number of parameters to 4.02M, significantly enhancing the multi-scale feature fusion capability and optimizing the model structure; PSWTConv (Model B) and C3k2CARA (Model C) contribute +0.9% and +0.7% accuracy gains, respectively, validating their effectiveness in preserving details of small targets and focusing attention. The three technologies were integrated in a synergistic manner (model F, namely YOLO26s-SOD) to achieve optimal performance: mAP@0.5 reached 49.1% (+9.2%), with only 4.32M parameters and 57.4 GFLOPs, achieving a good balance between accuracy and lightweight design.
[0088] Table 1. Ablation test results:
[0089]
[0090] Further comparison with different scale models of the YOLO26 series is shown in Table 2. YOLO26s-SOD, with less than half the number of parameters of YOLO26s (4.32M vs. 9.47M), surpasses all original variants, including YOLO26x (47.0%). Moreover, its GFLOPs (57.4) are much lower than those of YOLO26m (67.9), YOLO26l (86.1) and YOLO26x (193.4), fully meeting the dual requirements of high precision and low computational overhead for UAV platforms.
[0091] Table 2. Comparison results of YOLO26 models of different sizes and the improved model:
[0092]
[0093] The fine-grained evaluation on the VisDrone-DET2019 dataset, shown in Table 3, further confirms the model's advantages. On the validation set, YOLO26s-SOD significantly outperformed typical small target categories such as pedestrians (+13.8%), bicycles (+7.2%), tricycles (+8.4%), and buses (+11.2%), with the overall mAP@0.5 improving from 39.9% to 49.1%. On the more challenging test set, it maintained robust improvement, such as for pedestrians (32.0% → 42.6%), cars (72.5% → 80.0%), and tricycles (15.9% → 22.1%), with the overall mAP@0.5 reaching 39.0%, an improvement of 7.6 percentage points from the baseline. The results demonstrate that YOLO26s-SOD possesses stronger recognition capabilities, generalization, and robustness for various small targets in complex UAV aerial photography scenarios.
[0094] Table 3 Comparison of detection accuracy for each category in the VisDrone-DET2019 dataset:
[0095]
[0096] In such Figure 7 The images shown represent three typical scenes: a nighttime intersection, a farmland highway, and a daytime street. We use a three-color bounding box visualization technique to measure the improvement of this application: red indicates no target detection, blue indicates target detection but incorrect category, and green indicates correct detection. In the nighttime image, YOLO26s-SOD shows significant effectiveness in detecting occluded vehicles in the parking lot in the lower left corner and can detect small, distant vehicles. For the ultra-high-angle farmland highway scene, the baseline model YOLO26s performs poorly, detecting only a few targets but failing to identify any correct object categories. The improved model YOLO26s-SOD significantly improves this, successfully detecting most targets and correctly identifying half of the categories. In the typical daytime street scene with normal height and ample lighting, the improved model improves the overall target detection accuracy across the entire image, performing excellently for both near and far objects, as well as occluded objects.
[0097] In summary, this application improves the YOLO26s model by introducing a windmill-shaped wavelet transform convolution module in the downsampling stage of the backbone network, which integrates wavelet transform and multi-directional windmill-shaped convolution to simultaneously extract frequency domain details and spatial context information.
[0098] The A2C2f module and DySample dynamic upsampling are introduced into the neck network to optimize the multi-scale feature fusion path and enhance the semantic expression of shallow high-resolution features.
[0099] By embedding a context anchor residual attention feature fusion module in the backbone network and the neck network, the dual-branch context anchor attention based on the iRMB residual structure and CAA evolution is integrated into the C3k2 architecture to form a configurable hierarchical attention mechanism, which significantly improves the spatial focusing ability of small targets.
[0100] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for detecting small targets on unmanned aerial vehicles (UAVs) based on an improved version of YOLO26s, characterized in that: A YOLO26s-SOD model for small target detection in UAVs is constructed. This YOLO26s-SOD model is an improved YOLO26s model. The YOLO26s-SOD model includes a backbone network for feature extraction of the input image, a neck network for feature extraction and fusion of the feature map, and a head structure for detection and classification of the fused feature map output by the neck network. Skip connections are set between the backbone network and the neck network. Replace the first and second C3k2 modules of the original backbone network with context anchor residual attention feature fusion modules, replace the third C3k2 module with an A2C2f module, replace the convolution module between the context anchor residual attention feature fusion module and the corresponding A2C2f module with a windmill wavelet transform convolution module, delete the modules after the windmill wavelet transform convolution module and connect them to the first convolution module; The first Upsample module in the original neck network is replaced with the first connection module, the context anchor residual attention feature fusion module, and the DySample module connected in sequence. The second Upsample module is replaced with the DySample module. The C3k2 module in the neck network is replaced with the context anchor residual attention feature fusion module. After the second C3k2 module, the DySample module, the second connection module, and the context anchor residual attention feature fusion module are added in sequence. The images captured by the drone are input into the YOLO26s-SOD model, and the detection results are output.
2. The UAV small target detection method based on YOLO26s improvement as described in claim 1, characterized in that: The skip connection includes a second convolutional module, a third convolutional module, a fourth convolutional module, a fifth convolutional module, and a sixth convolutional module; The second convolutional module is used to perform a convolution operation on the features output by the first convolutional module of the backbone network and then feed them to the second connection module. The third convolutional module is used to perform a convolution operation on the features output by the first context anchor residual attention feature fusion module of the backbone network and then feed them to the second connection module of the neck network. The fourth convolutional module is used to perform a convolution operation on the features output by the first context anchor residual attention feature fusion module of the backbone network and then feed them to the first connection module of the neck network. The fifth convolutional module is used to perform a convolution operation on the features output by the second context anchor residual attention feature fusion module of the backbone network and then feed them to the first connection module. The sixth convolutional module is used to perform a convolution operation on the features output by the second context anchor residual attention feature fusion module of the neck network and then feed them to the last connection module of the neck network.
3. The UAV small target detection method based on YOLO26s improvement as described in claim 2, characterized in that: This windmill-shaped wavelet transform convolution module converts the feature map Input the residual connection branch and the channel expansion branch respectively; This residual connection branch will connect the feature map Input a 1×1 convolutional module, perform channel adjustment and downsampling, and output residual connection features; The channel expansion branch pairs with the feature map. Adjust the number of channels to obtain the feature map. A multi-level wavelet transform module is used to process the feature map. Processing: The first-level wavelet transform module transforms the feature map Decomposed into Sub-band Sub-band Sub-band and The four sub-bands are concatenated and then subjected to joint feature modulation via a channel-grouped 1×1 convolutional module. The output is also divided into four groups of enhanced sub-bands. , , and The four sets of enhanced subband input inverse wavelet transform modules are reconstructed back to the original spatial resolution to generate multi-scale fused feature maps. ; Expanding the branch also includes the feature map Input convolution module for feature extraction to obtain convolutional features ; Multi-scale fusion feature maps With convolutional features Add and fuse to generate wavelet-enhanced feature maps ; Four sets of depthwise convolutions were used to enhance the wavelet feature map. Multi-directional spatial feature extraction is performed, and the output features are added element-wise to complete the initial fusion of directional features. The fusion results are then sequentially processed through batch normalization and activation functions to generate directionally enhanced spatial features. ; The wavelet-enhanced feature map Spatial features Element-wise addition is performed, followed by 1×1 convolution for channel integration, and the enhanced features are output. The enhancement feature is added element-wise to the residual connection feature to output the first feature map.
4. The UAV small target detection method based on YOLO26s improvement as described in claim 3, characterized in that: The context anchor residual attention feature fusion module performs channel expansion on the input features to obtain expanded features. These expanded features are used for branch selection based on configuration parameters, with the specific rules as follows: When the configuration parameter attn=True, multiple SequentiaIList modules are used to process this extended feature. The SequentiaIList module includes the PSABlock-CARA module and the Bottleneck-CARA module connected in sequence. When the configuration parameter c3k=True, the C3k-CARA module is used to process this extended feature; When both configuration parameters c3k and attn are False, the extended feature is processed using Bottleneck-CARA. The extended feature is then concatenated with the features output from each branch, and the concatenated feature map is output via a convolution module.