A feature enhancement method for small target detection of unmanned aerial vehicle
By introducing CAB, MABlock, and SCM modules into the UAV aerial image detection model, the problem of low detection accuracy for small targets is solved, the detection accuracy and reliability are improved, and it is adapted to complex aerial photography scenarios.
Patent Information
- Application Number
- CN202610463566.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-09
- Publication Date
- 2026-07-03
Smart Images
Figure CN122336492A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of small target detection technology in UAV aerial imagery, and more specifically to a feature enhancement method for small target detection in UAVs. Background Technology
[0002] Currently, in UAV vision applications, the YOLO series has become the mainstream framework for real-time small target detection due to its excellent balance between speed and accuracy. YOLOv8n consists of three main parts: a backbone network for extracting multi-scale features, a neck network for fusing features from different levels, and a head for predicting target categories and bounding boxes.
[0003] However, when directly applied to drone aerial photography scenarios, these models face significant limitations: First, because small targets occupy very few pixels in an image, their deep features are easily lost due to reduced spatial resolution after multiple downsampling processes by the network; second, pre-set fixed-scale anchor boxes are difficult to adapt to the scale distribution of small targets in aerial images, leading to localization errors; third, existing multi-scale feature fusion mechanisms are not sensitive enough to small targets, making it difficult to effectively improve their detection capabilities. Furthermore, small targets themselves have sparse texture information and weak appearance features, resulting in feature representation capabilities far inferior to conventional targets. Images captured by drones typically feature dense targets, large scale differences, and complex background interference, further increasing the detection difficulty and easily leading to missed detections and false detections. These problems collectively affect the practical application performance of existing detectors in drone scenarios.
[0004] Therefore, designing a feature enhancement method for small target detection in UAVs to effectively improve the accuracy of small target detection, thereby enhancing the overall reliability and practicality, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of this, the present invention provides a feature enhancement method for small target detection in UAVs, which can effectively improve the detection accuracy of small targets and enhance the overall reliability and practicality.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: A feature enhancement method for small target detection in UAVs includes: The input image is fed into a detection network built on YOLOv8n; In the backbone of the detection network, the downsampling convolution of YOLOv8n is replaced by the CAB module. The CAB module adopts a multi-branch structure to fuse local details and contextual information, and combines the ECA attention mechanism to achieve channel-level adaptive calibration to obtain the initial feature map. In the backbone network and feature fusion layer of the detection network, the MABlock module is used to replace the C2f module of YOLOv8n. The MABlock module performs channel recombination, spatial modeling and multi-level nonlinear transformation on the input features of the initial feature map through multi-path parallel processing, and performs channel concatenation and fusion of the outputs of each path to generate the final feature map. In the feature fusion layer of the detection network, an SCM module is embedded after the feature concatenation operation. The SCM module adaptively refocuses the final feature map, suppresses background noise and aggregates the target response, and then passes the optimized features to the detection head. The detection head of the detection network outputs the category and location information of small target detections.
[0007] Preferably, the CAB module includes: detail branch, structure branch, and context branch; the detail branch is used to retain high-frequency information, the structure branch is used to extract lightweight structures, and the context branch is used to capture global dependencies by combining average pooling.
[0008] Preferably, the process of obtaining the initial feature map includes: concatenating the outputs of the detail branch, the structure branch and the context branch in the channel dimension, fusing them through convolution and then connecting them to the ECA attention mechanism; The ECA attention mechanism compresses the spatial dimension through global average pooling, then uses Sigmoid activation to generate weights and multiplies them element-wise with the original features to obtain the initial feature map.
[0009] Preferably, the MABlock module includes: a channel recombination unit, a spatial modeling unit, a direct mapping unit, and an n-layer nonlinear transformation unit; The channel reorganization unit linearly reorganizes the channels of intermediate features; the spatial modeling unit uses a bottleneck structure containing depthwise separable convolutions to efficiently capture local spatial context; the direct mapping unit splits the intermediate features in two along the channels, with one half directly retained as an identity mapping branch, and the other half sequentially undergoing a deep nonlinear transformation through n cascaded Bottleneck modules.
[0010] Preferably, the outputs of the channel recombination unit, spatial modeling unit, direct mapping unit, and n-layer nonlinear transformation unit are spliced and fused along the channel dimension to obtain the final feature map.
[0011] Preferably, the adaptive refocusing process of the SCM module on the final feature map includes: After receiving the final feature map, the SCM module generates a center heatmap using the Sigmoid activation function; A flow intensity map is generated using an additional Sigmoid activation function, and a two-dimensional flow field is predicted. Based on the two-dimensional flow field, the SCM module constructs an offset sampling grid using FWBlock. Bilinear interpolation is used to resample the final feature map to generate agglomerative feature maps; The enhanced output is obtained by adding the element-wise product of the flow intensity map, the condensation feature map, and the central thermal map to the final feature map. The enhanced output is then processed by convolution, group normalization, and SiLU activation function, and the optimized features are output through residual connection.
[0012] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a feature enhancement method for small target detection of UAVs, with the following beneficial effects: 1. In object detection tasks, while traditional downsampling operations can reduce spatial resolution, they easily lead to the loss of details and contextual information of small objects. To alleviate this problem, this patent designs a context-aware block (CAB). While performing spatial downsampling, the CAB collaboratively processes details, structure, and contextual information, better preserving details in shallow features and enhancing the initial feature response of small objects.
[0013] 2. To enhance the feature representation capability of the backbone network in complex scenarios, especially when dealing with the challenge of significant target scale variations in UAV aerial images, a hybrid aggregation block (MABlock) is introduced. This design achieves richer feature representation through fusion channel recombination, spatial modeling, and multi-level nonlinear transformations, which helps improve the network's adaptability to small targets and scale variations.
[0014] 3. Design a semantic cohesion module (SCM) to guide weak small target features to potential semantic centers, thereby enhancing the response strength to small targets and improving recognition ability. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0016] Figure 1 This is a schematic diagram of the CASDet structure provided by the present invention.
[0017] Figure 2 This is a diagram of the YOLOv8n framework provided by the present invention.
[0018] Figure 3 This is a CAB framework diagram provided for the present invention.
[0019] Figure 4 The SCM framework diagram provided by this invention. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] This invention proposes CASDet, a feature enhancement method based on YOLOv8n for small target detection in UAV scenarios, such as... Figure 1 As shown. The original YOLOv8n framework diagram is as follows. Figure 2 As shown. This invention is applicable to the detection of small targets in drone aerial images. Typical application scenarios include: urban traffic monitoring (such as pedestrian and vehicle detection) and public safety and infrastructure inspection (such as power line inspection and small target localization in emergency rescue).
[0022] The detection method described in this invention can be integrated into a UAV ground station or airborne processing system to perform small target detection on aerial images or video streams, and output target category and location information to support upper-level tasks such as security alarms, target tracking, or automated inspection report generation.
[0023] This invention discloses a feature enhancement method for small target detection on unmanned aerial vehicles (UAVs), comprising: The input image is fed into a detection network built on YOLOv8n; In the backbone of the detection network, the downsampling convolution of YOLOv8n is replaced by the CAB module. The CAB module adopts a multi-branch structure to fuse local details and contextual information, and combines the ECA attention mechanism to achieve channel-level adaptive calibration to obtain the initial feature map. In the backbone network and feature fusion layer of the detection network, the MABlock module is used to replace the C2f module of YOLOv8n. The MABlock module performs channel recombination, spatial modeling and multi-level nonlinear transformation on the input features of the initial feature map through multi-path parallel processing, and performs channel concatenation and fusion of the outputs of each path to generate the final feature map. In the feature fusion layer of the detection network, an SCM module is embedded after the feature concatenation operation. The SCM module adaptively refocuses the final feature map, suppresses background noise and aggregates the target response, and then passes the optimized features to the detection head. The detection head of the detection network outputs the category and location information of small target detections.
[0024] Specifically, the CAB module includes: detail branch, structure branch, and context branch; the detail branch is used to retain high-frequency information, the structure branch is used to extract lightweight structures, and the context branch is used to capture global dependencies by combining average pooling.
[0025] Specifically, the process of obtaining the initial feature map includes: concatenating the outputs of the detail branch, the structure branch, and the context branch in the channel dimension, fusing them through convolution, and then connecting them to the ECA attention mechanism; The ECA attention mechanism compresses the spatial dimension through global average pooling, then uses Sigmoid activation to generate weights and multiplies them element-wise with the original features to obtain the initial feature map.
[0026] In a specific embodiment of the present invention, to improve the network's response capability to small targets, this embodiment designs a CAB (Context-Aware Block) to replace the first four downsampled convolutions of the backbone network. For example... Figure 3 As shown, this module adopts a multi-branch structure to fuse local details and contextual information, and combines the ECA attention mechanism to achieve channel-level adaptive calibration, thereby enhancing the spatial discriminative ability of the initial features.
[0027] Specifically, the CAB module takes (B, cin, H, W) as input and outputs a channel number transformation to c. out Furthermore, the spatial dimensions are halved (B, cout, H / 2, W / 2), achieving a 2x downsampling function similar to the backbone's downsampling convolution. The module internally employs three parallel branches to extract multi-scale features: Detail branch 1 uses a double 3×3 concatenation with Skip connections (composed of 2×2 average pooling followed by 1×1 convolution) to retain high-frequency information; Structure branch 2 utilizes 1×1 followed by 3×3 convolutions for lightweight structure extraction; and Context branch 3 uses a 7×7 large convolution kernel combined with average pooling to capture global dependencies. The three outputs are concatenated and fused with 1×1 convolutions before being fed into the ECA attention mechanism. Specifically, the ECA attention mechanism compresses the spatial dimension through global average pooling, then uses a 1D convolution with a kernel size of 3 and a group number equal to the number of channels to perform local interactions between adjacent channels. Finally, it generates weights through Sigmoid activation and multiplies them element-wise with the original features, thus achieving efficient channel reweighting without significantly increasing the number of parameters.
[0028] Specifically, the MABlock module includes: a channel recombination unit, a spatial modeling unit, a direct mapping unit, and an n-layer nonlinear transformation unit; The channel recombination unit performs linear recombination of the channels of intermediate features; the spatial modeling unit uses a bottleneck structure containing depthwise separable convolutions to efficiently capture local spatial context; the direct mapping unit is used to split the intermediate features in two along the channels, with one half directly retained as an identity mapping branch, and the other half sequentially undergoing deep nonlinear transformations through n cascaded Bottleneck modules.
[0029] Specifically, the outputs of the channel recombination unit, spatial modeling unit, direct mapping unit, and n-layer nonlinear transformation unit are spliced and fused along the channel dimension to obtain the final feature map.
[0030] In a specific embodiment of the present invention, MABlock (Mixed Aggregation Block) is used instead of the standard C2f module and applied in the backbone and feature fusion layer. For example... Figure 1 As shown in the top left corner, MABlock enhances deep representations through multi-scale feature aggregation, alleviates the weakening of small target features during propagation, and improves the expressive power of multi-scale features.
[0031] Specifically, the overall structure of the MABlock module begins with an intermediate feature generated by a 1×1 convolution, which is then branched into multiple parallel paths to construct rich feature representations. In these paths, the first branch directly performs a linear recombination of the channels using a 1×1 convolution; the second branch employs a bottleneck structure containing depthwise separable convolutions to efficiently capture local spatial context; simultaneously, the intermediate feature is split in half along the channel, with one half directly retained as an identity mapping branch, and the other half sequentially passing through n cascaded Bottleneck modules. Each module contains two stacked 3×3 convolutions and residual connections, thus achieving deep nonlinear transformations.
[0032] Finally, the outputs of these 4+n different paths (including channel recombination, spatial modeling, direct mapping, and n-layer nonlinear transformation) are concatenated along the channel dimension and fused by a 1×1 convolution at the end to generate a final feature map that combines multi-scale perception and deep semantic information.
[0033] Specifically, the process of adaptive refocusing of the final feature map by the SCM module includes: After receiving the final feature map, the SCM module generates a center heatmap using the Sigmoid activation function; A flow intensity map is generated using an additional Sigmoid activation function, and a two-dimensional flow field is predicted. Based on the two-dimensional flow field, the SCM module constructs an offset sampling grid using FWBlock. Bilinear interpolation is used to resample the final feature map to generate agglomerative feature maps; The enhanced output is obtained by adding the element-wise product of the flow intensity map, the condensation feature map, and the central thermal map to the final feature map. The enhanced output is then processed by convolution, group normalization, and SiLU activation function, and the optimized features are output through residual connection.
[0034] In a specific embodiment of the present invention, such as Figure 4 As shown, the Semantic Condensation Module (SCM) is embedded in the feature fusion layer. It uses a learnable flow field to adaptively refocus the fused features, enhancing the feature response of the target region. After receiving the input feature map X, the SCM module performs three independent steps in parallel: First, it generates a center heatmap Mc using a 1×1 convolution with a sigmoid activation function, representing the confidence of each spatial location as a potential semantic center; second, it generates a flow intensity map S using another 1×1 convolution with a sigmoid function to suppress background noise and modulate the enhancement intensity; simultaneously, it predicts a two-dimensional flow field F (i.e., a learnable flow field) using independent 1×1 convolutions, which directly indicates the sampling offset required to achieve centripetal feature flow. Based on the flow field F, the module constructs an offset sampling grid using FWBlock: this grid is formed by element-wise addition of a normalized coordinate grid and a rearranged spatial offset grid. Then, it resamples the original feature X using bilinear interpolation (i.e., adaptive refocusing) to generate a condensed feature map Xc. Finally, the enhanced output is obtained by adding the element-wise product of S, Xc, and Mc to the original feature X. This result is then processed by a convolutional module containing 1×1 convolution, group normalization, and SiLU activation function, and output through residual connections.
[0035] Furthermore, the data stream of this invention follows YOLOv8n ( Figure 2 The standard path from Backbone to Neck to Head has been replaced and embedded in key nodes: Input: After the image enters the network, it first passes through the CAB module (which replaces the first four downsampling convolutions of the original backbone) to extract high-discrimination initial features while reducing the resolution.
[0036] Backbone: When feature maps are passed within the Backbone, the original C2f structure is completely replaced by the MABlock module, which is responsible for deep semantic channel reorganization and nonlinear space modeling.
[0037] Feature Fusion Layer (Neck): When the multi-scale features output by the Backbone enter the feature pyramid structure, part of the C2f structure is replaced by the MABlock module. After the last three feature concatenation operations, the SCM module is embedded. The SCM adaptively refocuses the concatenated fused features, suppresses background noise and aggregates the target response, and then passes the optimized features to subsequent layers or the detection head.
[0038] Furthermore, the present invention conducted ablation experiments on the three modules. As can be seen from Table 1, compared with YOLOv8n, each module can effectively improve the detection performance, whether used alone or in combination.
[0039] Table 1 shows the ablation experimental results on the VisDrone 2019-DET validation set.
[0040] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0041] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A feature enhancement method for small target detection of unmanned aerial vehicles, characterized in that, include: The input image is fed into a detection network built on YOLOv8n; In the backbone of the detection network, the downsampling convolution of YOLOv8n is replaced by the CAB module. The CAB module adopts a multi-branch structure to fuse local details and contextual information, and combines the ECA attention mechanism to achieve channel-level adaptive calibration to obtain the initial feature map. In the backbone network and feature fusion layer of the detection network, the MABlock module is used to replace the C2f module of YOLOv8n. The MABlock module performs channel recombination, spatial modeling and multi-level nonlinear transformation on the input features of the initial feature map through multi-path parallel processing, and performs channel concatenation and fusion of the outputs of each path to generate the final feature map. In the feature fusion layer of the detection network, an SCM module is embedded after the feature concatenation operation. The SCM module adaptively refocuses the final feature map, suppresses background noise and aggregates the target response, and then passes the optimized features to the detection head. The detection head of the detection network outputs the category and location information of small target detections. 2.The feature enhancement method for small target detection of UAVs according to claim 1, characterized in that, The CAB module includes: detail branch, structure branch, and context branch; the detail branch is used to retain high-frequency information, the structure branch is used to extract lightweight structures, and the context branch is used to capture global dependencies by combining average pooling.
3. The feature enhancement method for small target detection of unmanned aerial vehicles according to claim 2, characterized in that, The process of obtaining the initial feature map includes: concatenating the outputs of the detail branch, the structure branch and the context branch in the channel dimension, fusing them through convolution and then connecting them to the ECA attention mechanism; The ECA attention mechanism compresses the spatial dimension through global average pooling, then uses Sigmoid activation to generate weights and multiplies them element-wise with the original features to obtain the initial feature map.
4. The feature enhancement method for small target detection in unmanned aerial vehicles according to claim 1, characterized in that, The MABlock module includes: a channel recombination unit, a spatial modeling unit, a direct mapping unit, and an n-layer nonlinear transformation unit; The channel reorganization unit linearly reorganizes the channels of intermediate features; the spatial modeling unit uses a bottleneck structure containing depthwise separable convolutions to efficiently capture local spatial context; the direct mapping unit splits the intermediate features in two along the channels, with one half directly retained as an identity mapping branch, and the other half sequentially undergoing a deep nonlinear transformation through n cascaded Bottleneck modules.
5. The feature enhancement method for small target detection in unmanned aerial vehicles according to claim 4, characterized in that, The outputs of the channel recombination unit, spatial modeling unit, direct mapping unit, and n-layer nonlinear transformation unit are spliced and fused along the channel dimension to obtain the final feature map.
6. The feature enhancement method for small target detection in unmanned aerial vehicles according to claim 4, characterized in that, The adaptive refocusing process of the SCM module on the final feature map includes: After receiving the final feature map, the SCM module generates a center heatmap using the Sigmoid activation function; A flow intensity map is generated using an additional Sigmoid activation function, and a two-dimensional flow field is predicted. Based on the two-dimensional flow field, the SCM module constructs an offset sampling grid using FWBlock. Bilinear interpolation is used to resample the final feature map to generate agglomerative feature maps; The enhanced output is obtained by adding the element-wise product of the flow intensity map, the condensation feature map, and the central thermal map to the final feature map. The enhanced output is then processed by convolution, group normalization, and SiLU activation function, and the optimized features are output through residual connection.