Unmanned aerial vehicle visual angle target detection method based on deep learning

Abstract

The invention discloses an unmanned aerial vehicle visual angle target detection method based on deep learning, and the method employs a YOLOv11-based improved target detection model to carry out the target detection of an input image, and comprises a data set construction and model training stage and a model reasoning stage: S10, the data set construction and model training stage; s20, a model reasoning stage: obtaining a to-be-detected image and executing preprocessing, wherein the preprocessing comprises one or more of size normalization and pixel normalization; inputting the preprocessed image into the trained improved target detection model for feature extraction and multi-scale feature fusion to obtain at least three layers of fused feature maps with different scales; performing classification prediction and bounding box regression prediction based on the fused feature map, and outputting a candidate detection frame and category confidence thereof; and performing post-processing on the candidate detection frame to obtain a final detection result, wherein the post-processing at least comprises non-maximum suppression and / or threshold screening.

Description

Deep learning-based UAV-view target detection method Technical Field

[0001] This invention belongs to the field of computer vision and deep learning technology, specifically relating to a deep learning-based method for target detection from the perspective of a drone. Background Technology

[0002] With the rapid development of drone technology, drones, due to their advantages such as high mobility, flexible deployment, and wide field of view, are widely used in fields such as security patrol, traffic monitoring, urban management, disaster assessment, and target search. In these application scenarios, target detection technology based on drone aerial imagery is a key foundation for achieving automatic target perception and intelligent analysis.

[0003] However, since drones are usually shot from high altitudes or at an angle, the targets in aerial images are often small in scale, densely distributed, varied in posture, and have complex backgrounds. In particular, targets such as pedestrians and vehicles occupy a limited pixel area in the image, which poses a significant challenge to the accuracy and stability of target detection algorithms.

[0004] Existing object detection methods are mostly based on convolutional neural networks, whose network structures are typically designed for images from a general viewpoint. When directly applied to drone aerial photography scenarios, they are prone to problems such as insufficient representation of small target features, inadequate fusion of multi-scale features, and weakening of target edge information, resulting in high false negative and false positive rates.

[0005] To address these issues, existing technologies have attempted to improve detection performance by increasing network depth, introducing multi-scale feature pyramids, or designing complex attention mechanisms. However, such methods often involve a significant increase in the number of model parameters and computational load, which is not conducive to deployment and application in computing-constrained drone platforms or real-time detection scenarios. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention provides a deep learning-based UAV perspective target detection method to improve the accuracy and stability of target detection in UAV aerial photography scenarios while taking into account both detection accuracy and inference efficiency.

[0007] To solve the above technical problems, the present invention adopts the following technical solution: a drone-based target detection method based on deep learning, which uses an improved target detection model based on YOLOv11 to detect targets in input images, including a dataset construction and model training stage and a model inference stage. S10, Dataset Construction and Model Training Stage: Acquire drone aerial images or video frames to form raw samples; Label the raw samples to obtain label files, the label information including target category and bounding box position information; Divide the labeled samples into training set, validation set, and test set according to a preset ratio; Perform data augmentation processing on the training set, the data augmentation including geometric transformation, illumination / color transformation, and combined augmentation. One or more of the following: The preprocessed and enhanced training set is input into the improved object detection model for training to obtain the trained model parameters; S20, Model Inference Stage: The image to be detected is acquired and preprocessed, the preprocessing including one or more of size normalization and pixel normalization; The preprocessed image is input into the trained improved object detection model for feature extraction and multi-scale feature fusion to obtain at least three layers of fused feature maps at different scales; Based on the fused feature maps, classification prediction and bounding box regression prediction are performed respectively, and candidate detection boxes and their class confidence scores are output; Post-processing is performed on the candidate detection boxes to obtain the final detection result, the post-processing including at least non-maximum suppression and / or threshold filtering.

[0008] In one possible implementation, the improved target detection model includes a backbone network, a neck network, and a detection head network, wherein the backbone network is used to extract multi-level features from the input image, the neck network is used to perform multi-scale fusion of the multi-level features, and the detection head network is used to output classification predictions and bounding box regression predictions at each scale.

[0009] In one possible implementation, the multi-scale feature map output by the improved target detection model includes at least three scale features, P3, P4, and P5, from the neck network, where P3 is used to enhance small target detection, and P4 and P5 are used to take into account medium and large target detection.

[0010] In one possible implementation, at least a portion of the structure in the backbone network used for local feature extraction includes a Feature Complementary Attention Module (FCAM). The FCAM includes: a Feature Complementary Mapping Unit (FCM) for generating complementary features from input features; an Attention Guidance Generation Unit (ADG) for generating guidance weights based on the complementary features; a Guided Heterogeneous Attention Unit (BGA) for performing joint attention modeling of the input features in spatial and semantic dimensions based on the guidance weights; and fusing the attention features with the input features using a residual method to output enhanced features.

[0011] In one possible implementation, the Feature Complementary Mapping Unit (FCM) further includes: segmenting the input features along the channel dimension to form semantic branch features and spatial branch features; performing different forms of feature transformation on the semantic branch features and spatial branch features respectively to obtain semantically enhanced features and spatially preserved features; performing complementary mapping and fusion on the semantically enhanced features and spatially preserved features based on channel weights and spatial weights, and outputting complementary features that simultaneously contain spatial positioning information and semantic discriminative information.

[0012] In one possible implementation, the windmill-shaped convolution PConv in the C3K2-PW module is a directional enhancement convolution structure, comprising: performing asymmetric padding on the input features in different directions to form padding features with at least two directional branches; performing convolution operations on each directional branch to extract response features in the corresponding direction, so that the effective receptive field of each branch presents a spatial directional bias; fusing the response features of each directional branch to obtain output features, wherein the fusion includes one or more of element-wise addition, channel concatenation followed by convolution mapping, or weighted fusion; thereby enabling the output features to form a windmill-shaped / spiral-arm-shaped directional combination receptive field to improve the representation ability of edge information of slender, small-scale and dense targets.

[0013] In one possible implementation, the C3K2-PW module is built based on the C3K2 structure, and in its Bottleneck unit, the PConv is used to replace the original 3×3 standard convolution, while maintaining the residual connection and feature aggregation method of Bottleneck, so as to enhance the orientation-sensitive local feature extraction capability while controlling the number of parameters and computational cost.

[0014] In one possible implementation, the Feature Enhancement Block (FEB) is a parallel multi-branch feature enhancement structure, including at least three parallel convolutional branches. Each convolutional branch performs convolution operations on the same input feature using different kernel shapes and / or different kernel sizes to extract feature responses with different receptive field ranges and directional characteristics. The output features of each convolutional branch are fused in the channel dimension, and convolution mapping is performed on the fused features to generate enhanced feature outputs. Through the parallel multi-branch convolution and fusion operations, the enhanced features simultaneously contain local detail information and contextual semantic information, thereby improving the discriminability and stability of small targets in the multi-scale feature fusion process.

[0015] The present invention has the following beneficial effects: (1) Improved small target detection performance: By introducing the complementary modeling and guided attention mechanism of FCAM into the backbone network, the effective feature expression of small-scale targets and edge details in UAV aerial images is enhanced, and the false detection and missed detection in complex backgrounds are reduced.

[0016] (2) Enhanced orientation-sensitive features: By introducing the C3K2-PW module into the neck network and replacing the standard 3×3 convolution with windmill-shaped convolution PConv in the Bottleneck, the fused features have a receptive field of orientation combination, which improves the ability to represent slender structures, dense target contours and orientation textures.

[0017] (3) Multi-scale detail and context joint enhancement: By setting up a parallel multi-branch feature enhancement structure of FEB in the neck network, the response features of different receptive fields are fused to improve the discriminability and stability of small targets in the multi-scale feature fusion process.

[0018] (4) Lightweight and deployment friendly: While enhancing the feature expression capability, it maintains the dimensional consistency of module replacement and the residual / aggregation framework, which is easy to integrate directly into the YOLOv11 structure and control the number of parameters and computation, making it suitable for deployment on UAV platforms with limited computing power.

[0019] (5) Enhanced generalization ability: By constructing a drone aerial photography dataset covering multiple altitudes, multiple perspectives and multiple environments, and combining it with data augmentation and scale balancing strategies, the model's adaptability to different scenarios and target scale changes is improved. Attached Figure Description

[0020] Figure 1 is a flowchart of a deep learning-based UAV perspective target detection method according to an embodiment of the present invention; Figure 2 is a schematic diagram of the structure of the windmill-shaped convolution PConv in a specific application example. Detailed Implementation

[0021] 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, not all, of the embodiments of the present invention. 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.

[0022] In drone aerial photography datasets, targets are typically distributed in small-scale forms across a large area of ​​the scene, and multiple target instances with significant scale differences often exist simultaneously in the same image. Processing this drone aerial photography data reveals that traditional detection models are prone to issues such as weakened response to small-scale targets, blurred target edges, and interference between features of adjacent targets during multiple downsampling and feature fusion processes.

[0023] After processing the above data using the target detection method from the UAV perspective described in this invention, the model can more fully retain the spatial location information and local detail features of small targets during the feature extraction and fusion stages. Specifically, in aerial images, the responses of small-scale targets in the feature map are more concentrated, the target contours and edge structures are clearer, and the feature discrimination between different targets is enhanced in densely distributed scenes.

[0024] Referring to Figure 1, a flowchart of a deep learning-based UAV perspective target detection method according to an embodiment of the present invention is shown. The method uses an improved target detection model based on YOLOv11 to detect targets in the input image, including a dataset construction and model training stage and a model inference stage. S10, Dataset Construction and Model Training Stage: Acquire UAV aerial images or video frames to form raw samples; Label the raw samples to obtain label files, with label information including target category and bounding box position information; Divide the labeled samples into training set, validation set, and test set according to a preset ratio; Perform data augmentation processing on the training set, including one or more of geometric transformation, illumination / color transformation, and combined augmentation; Input the preprocessed and augmented training set into the improved target detection model for training to obtain trained model parameters. S20, Model Inference Stage: Acquire the image to be detected and perform preprocessing, including one or more of size normalization and pixel normalization; Size normalization can scale the input image to the size required by the network, and pixel normalization can linearly scale pixel values ​​to stabilize training and inference. The preprocessed image is input into the trained improved object detection model for feature extraction and multi-scale feature fusion to obtain at least three layers of fused feature maps at different scales; classification prediction and bounding box regression prediction are performed based on the fused feature maps to output candidate detection boxes and their class confidence scores; post-processing is performed on the candidate detection boxes to obtain the final detection results, and the post-processing includes at least non-maximum suppression and / or threshold filtering.

[0025] The above-described deep learning-based UAV target detection method addresses the characteristics of small-scale, densely distributed targets and complex backgrounds in UAV aerial photography and other scenarios. By structurally improving the backbone and neck networks, it enhances the feature representation capabilities of small targets while balancing model lightweighting and inference efficiency.

[0026] In one possible implementation, the improved target detection model includes a backbone network, a neck network, and a detection head network, wherein the backbone network is used to extract multi-level features from the input image, the neck network is used to perform multi-scale fusion of the multi-level features, and the detection head network is used to output classification predictions and bounding box regression predictions at each scale.

[0027] In one specific application example, the multi-scale feature map output by the improved object detection model includes at least three scale features from the neck network: P3, P4, and P5. P3 is used to enhance the detection of small objects, while P4 and P5 are used to take into account the detection of medium and large objects.

[0028] In one application example, to enhance the discrimination and localization capabilities of small targets, at least part of the structure in the backbone network used for local feature extraction includes a Feature Complementary Attention Module (FCAM). FCAM constructs complementary representations of spatial and semantic information and enhances key regions under attention guidance, thereby improving the response strength to small-scale targets and their edge details with lower additional computational overhead. The Feature Complementary Attention Module (FCAM) ​​includes: a Feature Complementary Mapping Unit (FCM) for generating complementary features from input features; an attention-guided generation unit for generating guiding weights based on the complementary features; a guided heterogeneous attention unit for performing joint attention modeling of the spatial and semantic dimensions of the input features based on the guiding weights; and fusing the attention features with the input features through residual fusion to output enhanced features. Through this residual fusion, the stable expression of the original features can be preserved while enhancing target-related information, reducing the risk of gradient vanishing during training and improving convergence stability.

[0029] In a specific application example, the input feature is denoted as F, and its size can be C×H×W. First, FCM constructs a complementary representation of the input feature F, outputting a complementary feature F_comp. Then, the attention-guided generation module generates a guiding weight G_a using F_comp as input. Next, the guided heterogeneous attention module performs joint attention modeling of the spatial and semantic dimensions of the input feature F based on the guiding weight G_a, obtaining the attention feature F_att. Finally, the attention feature F_att is fused with the input feature F through a residual method to output the enhanced feature F'.

[0030] The residual fusion can be implemented using element-wise addition, i.e., F' = F_att + F; alternatively, linear mapping, normalization, or activation operators can be set before and after addition to improve training stability and expressive power. Preserving the basic information of the input features through the residual path avoids feature degradation caused by excessive attention suppression and is beneficial for gradient propagation in deep networks.

[0031] The attention-guided generation module generates guiding weights G_a based on complementary features F_comp, highlighting feature dimensions sensitive to small targets in subsequent attention calculations. As an optional implementation, the module can perform global statistics and mapping on F_comp, such as globally converging spatial dimensions to obtain a compact description, then concatenating and mapping with a multilayer perceptron / fully connected layer, and outputting guiding weights through a nonlinear function. The guiding weights can be channel weight vectors, spatial weight maps, or combinations thereof.

[0032] The guided heterogeneous attention module is used to perform joint attention modeling on the input feature F based on the guiding weight G_a. As an optional implementation, the guided heterogeneous attention module can generate a query Q, a key K, and a value V from the input feature F, and introduce the guiding weight G_a during the query or attention weight calculation process to guide the modulation of Q or guide the bias of attention allocation, thereby making the attention more focused on the region where small targets may appear and their related semantic channels, thus obtaining the attention feature F_att.

[0033] In one specific embodiment, FCM is used to achieve complementary fusion of spatial and semantic information with low computational overhead. The Feature Complementary Mapping Unit (FCM) further includes: segmenting the input features along the channel dimension to form semantic branch features and spatial branch features; performing different forms of feature transformations on the semantic branch features and spatial branch features respectively to obtain semantically enhanced features and spatially preserved features; performing complementary mapping and fusion on the semantically enhanced features and spatially preserved features based on channel weights and spatial weights, and outputting a complementary feature F_comp that simultaneously contains spatial positioning information and semantic discriminative information.

[0034] In a specific application example, the spatial branch can use a lighter convolutional mapping to preserve spatial details (e.g., 1×1 convolution or equivalent linear mapping) to retain information such as target location, edges and local textures as much as possible; the semantic branch can use convolutional mapping with stronger local modeling capabilities to enhance contextual semantics (e.g., 3×3 convolution or equivalent transformation) to improve the ability to distinguish between targets and non-targets in complex backgrounds.

[0035] Furthermore, channel weights and spatial weights can be used to measure the complementary contributions of semantic and spatial branches, respectively. As an optional implementation, channel weights can be obtained based on global convergence and nonlinear mapping to emphasize channels that contribute to category discrimination; spatial weights can be obtained based on local / global spatial statistics to emphasize regions where the target may appear. Complementary mapping and fusion can employ one or more of the following: element-wise addition, convolutional mapping after channel concatenation, or weighted fusion, to generate complementary features F_comp that possess both spatial localization and semantic discrimination capabilities.

[0036] When deploying a network, FCAM can be inserted as needed at different levels of the backbone network. For example, it can replace the original local feature extraction structure in the high-resolution feature layer or the small target sensitive layer to enhance the detailed response to small-scale targets. At the same time, complementary fusion and guided attention are introduced in the semantic layer to improve robustness in complex backgrounds, thereby improving the accuracy and stability of small target detection.

[0037] In a specific application example, to enhance the ability to capture the directional texture and fine-grained structure of small targets during the multi-scale feature fusion stage, the windmill-shaped convolution PConv in the C3K2-PW module is a directional enhancement convolution structure, which includes: performing asymmetric padding on the input features in different directions to form padding features with at least two directional branches; performing convolution operations on each directional branch to extract the response features of the corresponding direction, so that the effective receptive field of each branch presents a directional bias in space; fusing the response features of each directional branch to obtain the output features, wherein the fusion includes one or more of element-wise addition, channel concatenation followed by convolution mapping, or weighted fusion; thereby forming a windmill-shaped / spiral-shaped directional combination receptive field of the output features to enhance the representation ability of edge information of slender, small-scale and dense targets.

[0038] In one possible implementation, the C3K2-PW module is built based on the C3K2 structure, and the PConv is used to replace the original 3×3 standard convolution in its Bottleneck unit. This is to obtain a combined receptive field of orientation with a smaller parameter overhead and enhance the responsiveness of small target edges and slender structures. At the same time, the residual connection and feature aggregation method of Bottleneck are maintained to enhance the orientation-sensitive local feature extraction capability while controlling the number of parameters and computational cost.

[0039] In a specific application example, let the input feature be X. PConv uses a multi-directional branch + asymmetric padding + asymmetric convolution kernel to perform directional enhancement convolution calculation: First, asymmetric padding is performed on the input feature X in different directions to obtain padding features with at least two, preferably four, directional branches; where, asymmetric padding can be represented by (P(l,r,t,b)), which corresponds to the number of padding pixels in the left, right, up, and down directions, respectively.

[0040] Furthermore, in the preferred four-branch implementation, the first layer of parallel convolution includes two types of asymmetric convolution kernels: a horizontal 1×3 convolution kernel and a vertical 3×1 convolution kernel; and by using asymmetric padding in different directions, the effective receptive field of each branch is directionally offset, thereby forming a combined sampling pattern of "windmill-shaped / spiral-shaped" outward diffusion. For example, the four branch outputs can be set as follows: Branch 1: After performing asymmetric padding (P(1,0,0,3)) on X, convolution is performed using a 1×3 convolution kernel; Branch 2: After performing asymmetric padding (P(0,3,0,1)) on X, convolution is performed using a 3×1 convolution kernel; Branch 3: After performing asymmetric padding (P(0,1,3,0)) on X, convolution is performed using a 1×3 convolution kernel; Branch 4: After performing asymmetric padding (P(3,0,1,0)) on X, convolution is performed using a 3×1 convolution kernel.

[0041] To improve training stability and convergence speed, normalization and nonlinear activation operators can be set sequentially after each branch convolution; for example, BN and SiLU activation can be concatenated after each branch convolution to obtain the directional response features (X_1,X_2,X_3,X_4) of each branch.

[0042] In the branch fusion stage, the response features of branches in each direction can be fused to obtain the output features. Fusion methods can include: element-wise addition fusion (Y=X_1+X_2+X_3+X_4); convolutional mapping after channel-dimensional concatenation (Y=Conv(Concat(X_1,X_2,X_3,X_4))); and weighted fusion: setting learnable weights for each branch or generating weights based on statistics and then summing them by weight. Fusion integrates contextual information from different directions and different effective receptive fields, thereby enhancing the representation of dense small targets, slender edges, and directional textures.

[0043] In a specific application example, in addition to local directional feature extraction, PConv's four branches can also introduce directional modeling with longer-range dependencies: for example, expanding / rearranging the input features in the horizontal or vertical direction to form a sequence-like structure, applying asymmetric convolutions in the corresponding directions to the expanded features to capture long-range context, and then restoring the spatial layout through inverse transformation; then merging with other branches to take into account both local details and global context.

[0044] In an implementation more tightly coupled with the detection network, the fused orientation enhancement features can be further normalized to generate a spatially aware attention map, which is then multiplied element-wise with the input features to highlight key region responses and suppress background redundancy. This attention enhancement output can serve as the main branch output of Bottleneck and together with the residual branch of Bottleneck, it constitutes the final output, thereby enhancing the features related to small targets while maintaining gradient propagation stability.

[0045] The Bottleneck unit in the C3K2-PWD module adopts the structure shown in Figure 2: the input features are first channel-mapped by a 1×1 convolution, then enter the windmill-shaped convolution PConv to obtain directional enhancement features, and then channel recovery is performed by a 1×1 convolution; at the same time, the input features are element-wise added to the main branch output through the residual branch to obtain the output features of Bottleneck.

[0046] During module replacement, the 3×3 standard convolutions within the Bottleneck in the original C3K2 are replaced with the PConv, while keeping the residual connections of the Bottleneck unchanged. The output feature space size and number of channels of the PConv are consistent with the output of the replaced 3×3 standard convolutions to ensure that the input and output dimensions of the Bottleneck are consistent before and after the replacement, which facilitates direct replacement in the model structure without additional size alignment operations.

[0047] Furthermore, the C3K2-PWD module maintains the feature aggregation method of C3K2: it concatenates the transform branch outputs and bypass branch features of several Bottleneck units in the channel dimension, and performs channel compression and information recombination through convolutional mapping to generate the module output features. By introducing the directional bias receptive field of PConv within the Bottleneck, the module can enhance the ability to extract direction-sensitive local features while controlling the number of parameters and computational cost.

[0048] In network applications, the C3K2-PWD module can replace at least some of the C3K2 positions used for feature fusion in the neck network, for example, by placing it at the fusion node after upsampling and feature stitching. This allows the fused features to obtain stronger edge, contour, and orientation texture representation before entering the detection head, thereby improving the localization stability and detection accuracy of dense small targets and slender structural targets.

[0049] In drone aerial photography scenarios with clear directions, including roads, building edges, and vehicle travel directions, the model's response to slender structures and directional targets is more stable after processing with the method of this invention. Compared to detection models using standard convolutional structures, this method can generate more directionally consistent feature responses in small-scale target regions with obvious directional textures, thereby reducing the probability of target edge breaks or local missed detections.

[0050] In a specific application embodiment, the Feature Enhancement Block (FEB) is a parallel multi-branch feature enhancement structure, including at least three parallel convolutional branches. Each convolutional branch performs convolution operations on the same input feature using different kernel shapes and / or different kernel sizes to extract feature responses with different receptive field ranges and directional characteristics. The output features of each convolutional branch are fused in the channel dimension, and convolution mapping is performed on the fused features to generate enhanced feature outputs. Through the parallel multi-branch convolution and fusion operations, the enhanced features simultaneously contain local detail information and contextual semantic information, thereby improving the discriminability and stability of small targets in the multi-scale feature fusion process.

[0051] In one specific embodiment, the Feature Enhancement Block (FEB) is disposed in at least one scale branch of the neck network to perform parallel multi-branch enhancement on the feature map from the backbone network or after upsampling / downsampling fusion, in order to compensate for the loss of small target detail information caused by deep downsampling and improve the discriminability of small targets in complex backgrounds. Let the input feature of the FEB be X, and its size be C×H×W.

[0052] FEB employs a parallel multi-branch structure, preferably including four parallel convolutional branches. Each convolutional branch performs convolution operations on the same input feature X with different kernel shapes or sizes to form feature responses with different receptive field ranges and directional characteristics. Specifically, the four convolutional branches include: a first branch: a convolutional block with a 1×1 kernel, performing lightweight channel / local mapping on the input feature to preserve detail information; a second branch: a convolutional block with a 1×3 kernel, used to extract local contextual responses in the horizontal direction; a third branch: a convolutional block with a 3×1 kernel, used to extract local contextual responses in the vertical direction; and a fourth branch: a convolutional block with a 3×3 kernel, used to extract more complete local neighborhood texture and edge information. The stride of each branch convolution is preferably 1, and appropriate padding is used to ensure that the spatial size of the branch output is consistent with the input feature X.

[0053] In a specific application example, each convolutional block can be composed of a convolutional layer, a normalization layer, and a nonlinear activation layer connected in series, for example, composed of Conv+BN+SiLU; alternatively, GN, ReLU, or other equivalent normalization / activation operators can be selected according to the deployment scenario to balance training stability and inference efficiency. The output features of each branch are denoted as X_1, X_2, X_3, and X_4, respectively, and X_i∈R^{C_i×H×W}.

[0054] In the branch fusion stage, the output features of each convolutional branch are fused along the channel dimension, preferably through channel concatenation: X_cat = Concat(X_1, X_2, X_3, X_4). By concatenating, multi-scale, multi-directional responses from different convolutional kernel shapes / sizes can be aggregated at the same spatial location, thereby preserving local details and supplementing directional contextual information.

[0055] To reorganize and compress the information of the fused features, FEB further performs convolutional mapping on X_cat to generate enhanced feature output. A 1×1 convolutional kernel and a stride of 1 are used to perform fusion mapping on the concatenated features to obtain the FEB output feature Y, which has a size of C'×H×W. Here, C' can be set to be the same as the input channel C or adjusted according to the network configuration so as to match with subsequent feature fusion nodes or detection head interfaces.

[0056] In one specific application example, the FEB output feature Y can be residually fused with the input feature X (e.g., element-wise addition or weighted addition) before outputting, so as to enhance the relevant response of small targets while retaining the stable representation of the original features; in another optional implementation, the FEB directly outputs the enhanced feature Y for use by subsequent Concat, upsampling / downsampling or detection prediction modules.

[0057] Through the above parallel multi-branch convolution and fusion method, the 1×1 branch is conducive to maintaining fine-grained information and reducing redundancy, the 1×3 and 3×1 branches are conducive to introducing directional context and expanding the effective receptive field, and the 3×3 branch is conducive to enhancing local texture and edge contour. After the features of each branch are fused and mapped, they form enhanced features that simultaneously contain local detail information and contextual semantic information, thereby improving the discriminability and stability of small targets in the multi-scale feature fusion process.

[0058] Compared to target detection models without the aforementioned structural improvements, the embodiments of the present invention can more fully preserve the edge information, directional texture, and local detail features of small-scale targets in complex backgrounds and densely distributed scenes, alleviate the information weakening problem of small targets in the multi-scale feature fusion process, and improve the stability and robustness of the detection results.

[0059] Meanwhile, by introducing directional enhancement PConv into the C3K2 structure and keeping the residual connection and feature aggregation methods unchanged, this implementation enhances the feature representation capability while avoiding a significant increase in the number of model parameters and computational complexity, balancing detection accuracy and inference efficiency. It is suitable for deployment in edge devices or real-time detection scenarios with limited computing power.

[0060] The deep learning-based UAV target detection method of this invention addresses the challenges of densely distributed targets and complex backgrounds, such as urban blocks, parking areas, and crowded scenes, where UAV aerial photography data often suffers from strong background texture interference and small distances between targets. After processing with the method of this invention, the model focuses more intently on key target regions during multi-scale feature fusion, suppresses background redundancy responses, and significantly reduces feature aliasing between targets. Experimental observations show that in the aforementioned scenarios, the detection results obtained using the method of this invention exhibit better stability across different scale feature layers, and the distinguishability of small targets in the fused features is improved, thereby effectively improving detection consistency in dense target scenes.

[0061] In practical deployments, the input image resolution, model size, and inference configuration can be adjusted according to specific application requirements to achieve a balance between detection accuracy and real-time performance. The improved target detection model is compatible with the existing YOLOv11 inference framework and has good engineering adaptability and system integration convenience.

[0062] It should be understood that the exemplary embodiments described herein are illustrative and not restrictive. Although one or more embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art will understand that various changes in form and detail may be made without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. A deep learning-based UAV-based target detection method, characterized in that, An improved object detection model based on YOLOv11 is used to detect objects in the input image, including the dataset construction and model training stage and the model inference stage. S10, dataset construction and model training stage: collect drone aerial images or video frames to form raw samples; The original samples are labeled to obtain a label file. The label information includes the target category and bounding box location information. The labeled samples are divided into training set, validation set and test set according to a preset ratio; Perform data augmentation on the training set, the data augmentation including one or more of geometric transformation, lighting / color transformation, and combined augmentation; The preprocessed and enhanced training set is input into the improved object detection model for training to obtain the trained model parameters; S2 0. Model inference stage: Acquire the image to be detected and perform preprocessing, the preprocessing including one or more of size normalization and pixel normalization; The preprocessed image is input into the trained improved object detection model for feature extraction and multi-scale feature fusion to obtain at least three layers of fused feature maps at different scales; classification prediction and bounding box regression prediction are performed based on the fused feature maps to output candidate detection boxes and their class confidence scores; post-processing is performed on the candidate detection boxes to obtain the final detection results, and the post-processing includes at least non-maximum suppression and / or threshold filtering.

2. The UAV-based target detection method based on deep learning as described in claim 1, characterized in that, The improved target detection model includes a backbone network, a neck network, and a detection head network. The backbone network is used to extract multi-level features from the input image, the neck network is used to perform multi-scale fusion of the multi-level features, and the detection head network is used to output classification predictions and bounding box regression predictions at each scale.

3. The UAV-based target detection method based on deep learning as described in claim 1, characterized in that, The improved target detection model outputs a multi-scale feature map that includes at least three scale features from the neck network: P3, P4, and P5. P3 is used to enhance the detection of small targets, while P4 and P5 are used to take into account the detection of medium and large targets.

4. The UAV-based target detection method based on deep learning as described in claim 1, characterized in that, At least a portion of the structure in the backbone network used for local feature extraction includes a Feature Complementary Attention Module (FCAM). The FCAM includes: a Feature Complementary Mapping Unit (FCM) for generating complementary features from input features; an Attention Guidance Generation Unit (ADG) for generating guidance weights based on the complementary features; a Guided Heterogeneous Attention Unit (BGA) for performing joint attention modeling of spatial and semantic dimensions on the input features based on the guidance weights; and fusing the attention features with the input features through residuals to output enhanced features.

5. The UAV-based target detection method based on deep learning as described in claim 4, characterized in that, The Feature Complementary Mapping (FCM) unit further includes: segmenting the input features along the channel dimension to form semantic branch features and spatial branch features; performing different forms of feature transformation on the semantic branch features and spatial branch features respectively to obtain semantically enhanced features and spatially preserved features; performing complementary mapping and fusion on the semantically enhanced features and spatially preserved features based on channel weights and spatial weights to output complementary features that simultaneously contain spatial positioning information and semantic discrimination information.

6. The UAV-based target detection method based on deep learning as described in claim 1, characterized in that, The windmill-shaped convolution PConv in the C3K2-PW module is a directional enhancement convolution structure, which includes: performing asymmetric padding on the input features in different directions to form padding features with at least two directional branches; performing convolution operations on each directional branch to extract response features in the corresponding direction, so that the effective receptive field of each branch presents a spatial directional bias; fusing the response features of each directional branch to obtain the output features, wherein the fusion includes one or more of element-wise addition, channel concatenation followed by convolution mapping, or weighted fusion; thereby forming a windmill-shaped / spiral-shaped directional combination receptive field for the output features to improve the representation ability of edge information of slender, small-scale and dense targets.

7. The UAV-based target detection method based on deep learning as described in claim 1, characterized in that, The C3K2-PW module is built on the C3K2 structure, and in its Bottleneck unit, the PConv is used to replace the original 3×3 standard convolution, while maintaining the residual connection and feature aggregation method of Bottleneck, so as to enhance the orientation-sensitive local feature extraction capability while controlling the number of parameters and computational cost.

8. The UAV-based target detection method based on deep learning as described in claim 1, characterized in that, The Feature Enhancement Block (FEB) is a parallel multi-branch feature enhancement structure, comprising at least three parallel convolutional branches. Each convolutional branch performs convolution operations on the same input feature using different kernel shapes and / or different kernel sizes to extract feature responses with different receptive field ranges and directional characteristics. The output features of each convolutional branch are fused in the channel dimension, and convolution mapping is performed on the fused features to generate enhanced feature outputs. Through the parallel multi-branch convolution and fusion operations, the enhanced features simultaneously contain local detail information and contextual semantic information, thereby improving the discriminability and stability of small targets in the multi-scale feature fusion process.

Images