A high-precision aerial small target detection method, system, device and storage medium based on YOLO

Abstract

This invention discloses a high-precision aerial small target detection method, system, device, and storage medium based on YOLO, relating to the fields of computer vision and image processing technology. The method includes: acquiring aerial image data, adjusting the images to a uniform size, and constructing a high-precision aerial small target detection network; inputting the aerial image data into the high-precision aerial small target detection network, and extracting multi-scale feature maps using an improved backbone network; enhancing and fusing the multi-scale feature maps using an improved neck network to obtain enhanced multi-scale feature maps; and using an introduced small target detection head to perform target classification and bounding box regression on the enhanced multi-scale feature maps, outputting detection results containing target location and category information. This invention can improve the accuracy and stability of aerial small target detection, reduce missed detections and false detections, and enhance automated perception capabilities in practical scenarios such as traffic, rescue, and inspection.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, and in particular to a high-precision aerial small target detection method, system, device and storage medium based on YOLO. Background Technology

[0002] With the rapid development and popularization of drone technology, image analysis based on drone aerial photography is playing an increasingly important role in civilian and military fields such as traffic flow monitoring, power line inspection, agricultural monitoring, rapid disaster response and rescue, and military reconnaissance. These application scenarios typically generate massive amounts of high-resolution image data, requiring the rapid and accurate identification of targets of interest, such as vehicles, pedestrians, ships, and buildings. However, target detection in aerial images faces a series of severe technical challenges, mainly manifested in the difficulty of small target detection: due to the high flight altitude of drones, targets occupy a very low pixel proportion in the image, usually less than 32×32 pixels, or even smaller, lacking sufficient texture, color, and shape information, making it difficult for traditional detectors to extract effective features; from the aerial perspective, targets (such as vehicles and crowds in parking lots) are often densely packed, with severe mutual occlusion. At the same time, due to the variation in drone flight altitude, targets of the same category will exhibit huge scale differences in the image, requiring the model to have extremely strong multi-scale adaptability; the background of aerial images usually contains a variety of complex features such as buildings, trees, shadows, and roads, and this background information is intertwined with target features, forming strong semantic noise and scene noise. Furthermore, changes in lighting conditions, weather conditions (such as fog and cloud cover), and noise from imaging sensors further exacerbate the difficulty of feature extraction.

[0003] Currently, mainstream general-purpose object detection models, such as Faster R-CNN, SSD, and the YOLO series, have achieved great success in object detection at conventional scales. However, their performance drops sharply when faced with the challenge of small object detection because the deep feature maps of deep convolutional neural networks (CNNs) have low resolution, and the detailed information of small objects is almost completely lost after multiple downsampling. Although shallow feature maps have high resolution, they lack semantic information and are difficult to distinguish between objects and background. Traditional feature pyramid networks (FPNs) fuse features through simple top-down paths and lateral connections, but this fusion method is difficult to accurately capture and convey the spatial and semantic features of small objects in scenarios where small object information is sparse and the background is complex, and it is also prone to introducing noise. Existing models are quite sensitive to noise and interference in input images and lack effective denoising mechanisms, which leads to a large number of missed detections and false detections in complex environments. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention provides a high-precision aerial small target detection method, system, device and storage medium based on YOLO.

[0005] Therefore, the technical problem solved by this invention is the low detection accuracy caused by the small size, dense distribution, and complex background of targets in UAV aerial images.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a high-precision aerial small target detection method based on YOLO, comprising: Aerial image data is acquired, the images are adjusted to a uniform size, and a high-precision aerial small target detection network is constructed. The network is based on the YOLO architecture and includes an improved backbone network, an improved neck network, and an improved small target detection head. Aerial image data is input into a high-precision aerial small target detection network, and multi-scale feature maps are extracted using an improved backbone network. An improved neck network is used to enhance and fuse multi-scale feature maps, resulting in enhanced multi-scale feature maps. The introduced small target detection head is used to perform target classification and bounding box regression on the enhanced multi-scale feature map, and the output is the detection result containing target location information and category information.

[0007] As a preferred solution for a high-precision aerial small target detection method based on YOLO, the following is provided: The process involves acquiring aerial image data, adjusting the images to a uniform size, and constructing a high-precision aerial small target detection network. This network is based on the YOLO architecture and includes an improved backbone network, an improved neck network, and an improved small target detection head. In the backbone and neck networks, the original C2f modules are replaced by C2f_DCWR modules improved based on multi-scale dilated convolution residual modules; in the backbone network, a multi-module collaborative denoising module is introduced; in the neck network, an improved multi-layer feature screening and fusion pyramid network is used to replace the original feature pyramid structure; and a small target detection head is introduced in the head network. The C2f_DCWR module, which is an improvement based on the multi-scale dilated convolution residual module, extracts multi-scale contextual information by constructing a bottleneck structure and adopting a two-step strategy. The small target detection head constructs the P2 detection head by utilizing the high-resolution feature map of the shallowest layer of the backbone network, based on the multi-layer feature screening and fusion pyramid network.

[0008] As a preferred solution for a high-precision aerial small target detection method based on YOLO, the following is provided: The step of inputting aerial image data into a high-precision aerial small target detection network and extracting multi-scale feature maps using an improved backbone network includes: Aerial image data is input into the backbone network. The input feature map is processed by multiple C2f_DCWR modules in sequence. Each C2f_DCWR module extracts regional residual features and performs standard convolution and activation processing on the input features to generate regional residual features. The regional residual features are divided into multiple subgroups in the channel dimension by multi-scale morphological filtering. Deep convolution with different dilation rates is applied in parallel to each subgroup to obtain multi-scale receptive field information. After feature integration, semantic residual features are formed. The semantic residual features are added element-wise to the initial input, and the processing result is output.

[0009] As a preferred solution for a high-precision aerial small target detection method based on YOLO, the following is provided: The step of inputting aerial image data into a high-precision aerial small target detection network and extracting multi-scale feature maps using an improved backbone network also includes: A multi-module collaborative denoising module is used to suppress and enhance noise in the features. After the noisy input features enter the multi-module collaborative denoising module, they first pass through sparse block processing. The sparse block forms a sparse distribution of features by alternating stacking of dilated convolution and standard convolution. Then, it passes through the feature enhancement block, which concatenates the output of the sparse block with the original input of the module and outputs the fused features after activation processing. Finally, it passes through the attention block, which generates attention weights through convolution operations. These weights are then multiplied element-wise with the output of the feature enhancement block to obtain the predicted noise map. The clear features are then reconstructed through the residual operation between the noisy input features and the noise map.

[0010] As a preferred solution for a high-precision aerial small target detection method based on YOLO, the following is provided: The improved neck network is used to enhance and fuse multi-scale feature maps to obtain enhanced multi-scale feature maps, including: The multi-scale feature maps output from the backbone network are input into the neck network. The neck network adopts a multi-layer feature selection and fusion pyramid structure. The feature selection module and the selective feature fusion module sequentially preprocess and enhance the features of each level. The feature selection module preprocesses each feature map layer. The channel attention submodule in the feature selection module extracts global information of each channel through parallel global pooling operations. After activation, it generates channel attention weights. The channel attention weights are multiplied element-wise with the original features to achieve the selection of key features. The feature alignment submodule in the feature selection module performs convolution operations on each feature map layer to compress the number of channels of each feature map layer to a preset dimension.

[0011] As a preferred solution for a high-precision aerial small target detection method based on YOLO, the following is provided: The method of using an improved neck network to enhance and fuse multi-scale feature maps to obtain enhanced multi-scale feature maps also includes: The selective feature fusion module performs top-down enhancement fusion of features at each level. For any set of high-level and low-level features to be fused, the high-level features are first upsampled to match the spatial resolution with the low-level features. Then, the upsampled high-level features are input into the channel attention submodule to generate semantic attention weights. The semantic attention weights are used to weight and filter the low-level features. Finally, the filtered low-level features are added element-wise to the upsampled high-level features to obtain the fused output feature map.

[0012] As a preferred solution for a high-precision aerial small target detection method based on YOLO, the following is provided: The method utilizes an introduced small target detection head to perform target classification and bounding box regression on the enhanced multi-scale feature map, outputting detection results containing target location and category information, including: The enhanced multi-scale feature map output by the neck network is input into the head network. The head network introduces an additional small target detection head on the basis of the original detection head, which is connected to the highest resolution enhanced feature map output by the neck network. The small target detection head directly predicts based on the highest resolution enhanced feature map. The original detection heads simultaneously perform target classification and bounding box regression on the enhanced feature maps of the corresponding scale. Each detection head adopts a decoupled structure and outputs target category information and bounding box position information independently. The outputs of all detection heads are fused and filtered through post-processing operations to remove redundant detection boxes, and the final output includes the detection results containing the target category label and bounding box coordinates.

[0013] Secondly, the present invention provides a high-precision aerial small target detection system based on YOLO, comprising: The data preprocessing and architecture building module is used to acquire aerial image data, adjust the images to a uniform size, and build a high-precision aerial small target detection network. The network is based on the YOLO architecture and includes an improved backbone network, an improved neck network, and an improved small target detection head. The multi-scale feature extraction module is used to input aerial image data into a high-precision aerial small target detection network and extract multi-scale feature maps using an improved backbone network. The feature enhancement and fusion module is used to enhance and fuse multi-scale feature maps using an improved neck network to obtain enhanced multi-scale feature maps. The small target detection output module is used to perform target classification and bounding box regression on the enhanced multi-scale feature map using the introduced small target detection head, and outputs detection results containing target location information and category information.

[0014] Thirdly, the present invention provides a computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the YOLO-based high-precision aerial small target detection method are implemented.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of a YOLO-based high-precision aerial small target detection method.

[0016] The beneficial effects of this invention are as follows: In UAV traffic monitoring scenarios, this invention can accurately identify small targets such as vehicles and pedestrians from complex high-altitude images. Even in situations with dense traffic, mutual obstruction, or shadow interference, it maintains stable detection capabilities, reducing missed and false alarms, and providing reliable data support for traffic flow statistics and violation evidence collection. At disaster relief sites, UAVs equipped with this method can quickly locate small targets such as trapped personnel and relief supplies, effectively overcoming visual interference from complex backgrounds such as smoke, dust, and building ruins, helping rescue personnel to grasp the situation on-site in a timely manner and improving emergency response efficiency. For long-distance inspection tasks such as power lines and oil and gas pipelines, this method can accurately detect minor hidden dangers such as insulator damage, equipment corrosion, and foreign object entanglement during high-speed UAV flight, reducing the workload of manual verification and improving the automation level of inspection operations. This method can also accurately identify subtle features such as early pest and disease spots and missing seedlings, assisting in precise pesticide application and replanting, reducing pesticide waste, and improving the precision of agricultural production management. In applications such as port monitoring and border patrol, this invention enables stable detection of small vessels and personnel from long-distance, wide-area aerial footage, maintaining a high detection success rate even in complex environments with varying lighting and water reflections, thus enhancing all-weather monitoring capabilities. By improving the detection accuracy and robustness of small aerial targets, this invention allows UAVs to more reliably complete automated target recognition tasks in various practical business scenarios, reducing the cost of manual intervention, improving the timeliness and accuracy of information acquisition, and providing a more solid visual perception foundation for various industry applications. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an overall flowchart of a high-precision aerial small target detection method based on YOLO provided by the present invention.

[0019] Figure 2 The diagram shows the overall network structure of the YOLO-based high-precision aerial small target detection method provided by this invention, in a simulation example.

[0020] Figure 3 The design diagram of the C2f_DCWR module, which introduces a DCWR bottleneck structure, is provided in the simulation example of a high-precision aerial small target detection method based on YOLO provided by this invention.

[0021] Figure 4 The network structure diagram of the multi-module collaborative denoising module (MMJD) in a simulation example of a high-precision aerial small target detection method based on YOLO provided by the present invention.

[0022] Figure 5 The diagram shows the structure of the multi-layer feature filtering fusion pyramid network (HLFPN) in a simulation example of a high-precision aerial small target detection method based on YOLO provided by this invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a high-precision aerial small target detection method based on YOLO, including: S1: Acquire aerial image data, adjust the images to a uniform size, and construct a high-precision aerial small target detection network. The network is based on the YOLO architecture and includes an improved backbone network, an improved neck network, and an improved small target detection head. S2: Input the aerial image data into a high-precision aerial small target detection network, and use the improved backbone network to extract multi-scale feature maps; S3: Use an improved neck network to enhance and fuse the multi-scale feature maps to obtain enhanced multi-scale feature maps; S4: Utilize the introduced small target detection head to perform target classification and bounding box regression on the enhanced multi-scale feature map, and output detection results containing target location information and category information.

[0025] It should be noted that through steps S1-S4, a multi-scale context extraction and noise suppression mechanism is introduced collaboratively in the backbone network and the neck network. The neck network achieves efficient screening and fusion of high-level semantics and shallow details, and a dedicated small target detection head is added to form an end-to-end detection process. Without increasing too much computational overhead, the detection accuracy and positioning accuracy of small targets in aerial images with complex backgrounds are significantly improved. It can stably output reliable target location and category information in practical application scenarios such as traffic monitoring, disaster relief, and security patrol, providing efficient technical support for UAV vision tasks.

[0026] Example 2, refer to Figure 1 As an embodiment of the present invention, based on the previous embodiment, a high-precision aerial small target detection method based on YOLO is provided, comprising: In this embodiment, aerial image data is acquired in step S1 above, the images are adjusted to a uniform size, and a high-precision aerial small target detection network is constructed. This network is based on the YOLO architecture and includes an improved backbone network, an improved neck network, and an improved small target detection head, comprising: The high-precision aerial small target detection model is built on the YOLO framework, including an improved backbone network, an improved neck network, and an improved small target detection head. In the backbone network and neck network, the original C2f module (CSPNet with two convolutions and feature map split, the basic feature extraction module in YOLOv8 network) is replaced by an improved C2f_DCWR module based on a multi-scale dilated convolution residual module.

[0027] In the backbone network, a multi-module collaborative denoising module (MMJD) is introduced; in the neck network, an improved multi-layer feature screening and fusion pyramid network is used to replace the original feature pyramid structure; and a small target detection head is introduced in the head network.

[0028] It should be noted that, to overcome the limitations of the original C2f module in multi-scale information extraction, this invention designs a novel bottleneck structure DCWR module (Dilated Convolutional Weighted Residual module), and uses it to construct the C2f_DCWR module. This module, through the design of the residual structure, ensures that gradients can be effectively backpropagated, alleviating the gradient vanishing problem in deep networks.

[0029] The DCWR module extracts multi-scale contextual information through a two-step strategy: Step 1: Region Residual Feature Extraction. The core objective of this step is to extract target-related residuals from the input features, generating an intermediate representation called region residual features, which provides a structured feature foundation for subsequent morphological filtering.

[0030] Specifically, the input feature map first undergoes a standard 3×3 convolution with an inverse bottleneck design, expanding the number of input channels from 2C to 3C to map the features to a higher-dimensional space and enhance representation capabilities. The specific value of C depends on the network layer. Subsequently, it is processed by batch normalization (BN) and the ReLU activation function to generate regional residual features containing target-related residuals. Here, the ReLU activation function suppresses background noise and enhances the edge response of small targets, achieving initial sparsity and refinement of the features, thereby improving the accuracy and efficiency of subsequent processing.

[0031] Step 2: Multi-scale morphological filtering. The core operation of this step is to use parallel multi-rate deep convolution to perform morphological filtering on regional features, aiming to extract key information from structures at different scales and ultimately output semantic residual features that carry deep semantics.

[0032] Specifically, the regional residual features output from the first step are divided into multiple subgroups along the channel dimension. For each subgroup, depthwise convolutions with different dilation rates are applied in parallel to obtain multi-scale receptive field information, ranging from local details to global semantics. Only a single desired receptive field is applied to each channel, thus avoiding redundant coverage caused by multiple receptive fields, improving computational efficiency, and reducing interference with fine-grained features. These parallel branches efficiently capture multi-scale context by performing morphological filtering on the regional feature maps with specific receptive fields. After extracting the multi-scale context information, the feature maps output from each branch are first concatenated at the channel level, then compressed and integrated through BN layers and pointwise convolutions to form the final semantic residual features. This design redefines the role of multi-rate dilated convolutions: it no longer focuses on capturing complex global semantics, but simplifies it to performing morphological filtering with specific receptive fields on concise regional feature maps, thereby more efficiently ensuring the acquisition of multi-scale context information.

[0033] Finally, the semantic residual feature is element-wise added to the initial input features of the DCWR module to obtain the output feature. This residual design allows the entire module to focus on learning the residual mapping between the input and output. Furthermore, considering that the ability of a large receptive field to establish effective spatial connections depends on the support of features from a small receptive field, in a preferred embodiment, the number of channels in the first branch (minimum dilation rate) is set to twice that of the other branches, thereby enhancing its ability to extract local features and effectively avoiding the feature dilution problem that may occur in small object detection.

[0034] Multi-module collaborative denoising module in the backbone network: To address the feature contamination problem caused by complex backgrounds and imaging noise, this invention introduces a multi-module collaborative denoising module into the backbone network. This module re-emphasizes the features of small targets obscured by noise, ultimately achieving more accurate and robust detection. The MMJD module employs an end-to-end residual learning framework, and its mathematical expression is: in: For noisy input features, For MMJD composite mapping functions, For the predicted noise mapping, To ensure clear reconstruction features, the MMJD module consists of three cascaded core sub-modules: Sparse Block (SB): This module differs structurally from traditional sparse mechanisms. Essentially, it's a lightweight network module built from dilated convolutions and standard convolutions. This module alternately stacks two basic units: dilated convolution + BN + ReLU and standard convolution + BN + ReLU, for a total of 12 layers. Dilated convolutions expand the receptive field to incorporate broader contextual information, while standard convolutions extract local features. All convolutional layers use 3×3 kernels. This alternating combination of high and low receptive fields naturally creates a sparse distribution of high-energy and low-energy points at the feature response level. This allows for focusing on key information with less computational cost, effectively controlling model complexity while improving denoising performance and training efficiency.

[0035] Feature Enhancement Block (FEB): To alleviate the problem of shallow feature decay caused by gradient vanishing in deep networks, this invention introduces a feature enhancement block in the multi-module collaborative denoising module. This module effectively enhances feature representation capabilities by constructing long-distance paths and fusing the global context and local details of the image. Specifically, the FEB concatenates the output of the sparse block with the original input (or shallow features) of the module along the channel dimension, expanding the number of output channels to 2C, forming a feature map of size 2C×H×W, where H represents the feature map height and W represents the feature map width. This operation effectively preserves the low-level structural information of the input image while fusing semantic features extracted from deeper layers. Finally, a nonlinear transformation is introduced through the Tanh activation function to further enhance the expressive power and numerical stability of the features.

[0036] Attention Block (AB): In image processing tasks, complex backgrounds often obscure key features, increasing the difficulty of model training. To address this, this invention introduces a lightweight attention block in the multi-module collaborative denoising module. This block aims to guide shallow feature learning using abstract information from deep features, effectively focusing on key noise regions. The module takes the fused features (2C channels) output from the previous stage (FEB) as input and compresses the high-dimensional features into an attention weight vector with the same number of channels as the original input image through a 1×1 convolutional layer with C output channels. Subsequently, the generated attention weights are multiplied element-wise with the original output features (C channels) of the FEB. This operation enhances significant noise regions and suppresses redundant background information, thus completing feature recalibration. This attention mechanism, through simple single-layer convolution and feature multiplication, achieves cross-layer feature guidance and adaptive noise perception without introducing complex computations, effectively improving denoising capabilities in complex backgrounds.

[0037] A multi-layer feature selection and fusion pyramid network in the neck network: To address the significant scale differences of targets in UAV aerial images due to varying shooting altitudes, this invention proposes a multi-layer feature selection and fusion pyramid network to replace the original YOLO feature pyramid, aiming to improve multi-scale feature fusion capabilities. The multi-layer feature selection and fusion pyramid network mainly consists of a feature selection module and a selective feature fusion module.

[0038] The feature selection module is used to preprocess features from different levels of the backbone network and consists of two core parts: the channel attention submodule and the feature alignment submodule.

[0039] The channel attention submodule is responsible for filtering key features. This module first performs max pooling and average pooling operations in parallel on the input feature map. Max pooling extracts the most representative and relevant information from each channel, preserving important features while ignoring irrelevant noise; average pooling, on the other hand, evenly summarizes all information in the feature map, avoiding excessive loss of important information and helping to maintain the integrity of the features. The pooled features are then processed through a sigmoid activation function to generate attention weights for each channel. By combining these two pooling methods, the channel attention submodule can extract the most representative key information while minimizing information loss. These weights are then multiplied element-wise with the original features to obtain the filtered feature map, which not only preserves the key regions in the original feature map but also effectively suppresses redundant or invalid features.

[0040] The feature alignment submodule is used to unify feature dimensions. Since feature maps at different scales originate from different layers in the network, their channel counts are usually inconsistent. Direct fusion would fail to achieve effective integration due to this dimensionality inconsistency. To address this issue, the feature alignment submodule applies a 1×1 convolution operation to each scale of the feature map. This operation not only compresses the channel dimension of the feature map but also quickly reconstructs and transforms channel information without changing the spatial resolution. In this way, the number of channels in feature maps across all scales is uniformly compressed to a preset 256 dimensions, thus ensuring structural consistency and computational feasibility in subsequent fusion operations.

[0041] The selective feature fusion module is located in the lateral connections of the multi-layer feature filtering and fusion pyramid network, and is used to efficiently fuse high-level semantic features and low-level detailed features. While high-level features contain rich semantic information, they have low spatial resolution and ambiguous localization; low-level features have high spatial resolution and precise localization, but limited semantic information. To effectively combine the advantages of both, the selective feature fusion module strategically fuses features, using high-level features as weights to filter key information from low-level features, thereby avoiding unnecessary feature redundancy.

[0042] First, for the input high-level feature maps, transposed convolution is used for upsampling to expand the spatial size of the feature maps, restore their resolution, and match the scale of the low-level feature maps. The parameters of transposed convolution are learnable, and its operation mechanism can naturally introduce dynamic receptive fields to different regions of the input, enabling more flexible modeling of irregular shapes and complex structures. While preserving rich semantic information, it significantly alleviates the problems of structural ambiguity and feature loss.

[0043] Then, the upsampled high-level feature map is input into the channel attention submodule to generate semantic attention weights. These weights contain the semantic orientation of the high-level features toward the key target.

[0044] Finally, the generated semantic attention weights are used to perform weighted filtering on the low-level feature maps, that is, retaining the key parts of the low-level features based on the semantic information contained in the high-level features. The filtered low-level feature maps are then fused with the upsampled high-level feature maps to obtain the final output feature map. This feature map integrates the semantic information of the high-level features and the precise spatial information of the low-level features, greatly enhancing the expressive power for small targets. Its fusion mechanism can be expressed by the formula: in, This represents the high-level feature map of the input. This represents the high-level feature map after transposed convolution upsampling. This represents the low-level feature map of the input. This is the function representation of the channel attention submodule. This represents the output feature map of the selective feature fusion module.

[0045] Small Target Detection Head in the Head Network: Based on the multi-layer feature filtering and fusion pyramid network, this invention introduces an additional P2 detection head for extremely small targets. The P2 layer utilizes high-resolution feature maps from the shallowest layer of the backbone network, which possess the highest spatial resolution and the richest detail information such as edges and textures. After enhanced fusion through the multi-layer feature filtering and fusion pyramid network, the P2 detection head can directly make predictions based on these high-resolution, semantically rich feature maps, thereby significantly improving the recall rate and localization accuracy for extremely small targets smaller than 32×32 pixels.

[0046] In this embodiment, step S2 above, which involves inputting aerial image data into a high-precision aerial small target detection network and extracting multi-scale feature maps using an improved backbone network, includes: Aerial image data first enters the backbone network, which integrates a C2f_DCWR module and a multi-module collaborative denoising module. Specifically, the input feature map is processed sequentially by multiple C2f_DCWR modules. Each C2f_DCWR module first performs a region residual feature extraction step, applying standard convolution, batch normalization, and ReLU activation to the input features to generate region residual features. Subsequently, a multi-scale morphological filtering step divides the region residual features into multiple subgroups along the channel dimension, applying depthwise convolutions with different dilation rates in parallel. After concatenation, batch normalization, and pointwise convolution, semantic residual features are formed. Finally, the semantic residual features are element-wise added to the module's initial input to output the module's processing result. Through the stacking of multiple C2f_DCWR modules, the backbone network progressively extracts features with multi-scale receptive fields at different levels.

[0047] Building upon this, a multi-module collaborative denoising module in the backbone network suppresses and enhances noise in the features. (Noisy input features) After entering the MMJD module, the first step is sparse block processing. The sparse block consists of 12 layers of alternating dilated and standard convolutions, creating a sparse distribution of features through alternating high and low receptive fields, focusing key information with low computational cost. Next, the feature enhancement block concatenates the output of the sparse block with the original input along the channel dimension, and outputs fused features after Tanh activation. Finally, the attention block generates attention weights through 1×1 convolutions, which are then multiplied element-wise with the output of the feature enhancement block to obtain the predicted noise map. And through residual calculation Reconstruct clear features .

[0048] After the above processing, the backbone network finally outputs multi-scale feature maps, including hierarchical features of different resolutions such as P2, P3, P4, and P5. The shallow feature maps retain high-resolution detail information, while the deep feature maps contain strong semantic information, providing input for further fusion of the neck network.

[0049] In this embodiment, step S3 above utilizes an improved neck network to enhance and fuse the multi-scale feature map, resulting in an enhanced multi-scale feature map including: Taking the P3, P4, and P5 layer features output by the backbone network as an example, preprocessing is first performed through the feature selection module. For each feature map, the channel attention submodule performs max pooling and average pooling in parallel, adds the pooling results, and generates channel attention weights through the Sigmoid function. These weights are then multiplied element-wise with the original features to filter key features. Subsequently, the feature alignment submodule applies 1×1 convolution to each feature map to uniformly compress the number of channels to a preset 256 dimensions, ensuring structural consistency in subsequent fusion operations.

[0050] After feature selection, top-down feature fusion is performed using a selective feature fusion module. Taking the fusion of high-level feature P5 and low-level feature P4 as an example: First, P5 is upsampled using a transposed convolution to match its spatial resolution with P4; the upsampled high-level feature map is then input into the channel attention submodule to generate semantic attention weights; these weights are used to weight and filter the low-level feature map P4; finally, the filtered low-level feature map is added element-wise to the upsampled high-level feature map to obtain the fused output feature map. Similarly, the fused feature map is fused with the shallower feature P3 using the same mechanism, forming a complete top-down feature enhancement path. Optionally, a bottom-up fusion path can also be constructed, where shallow features are downsampled and then fused again with high-level features using the selective feature fusion module to enhance the bidirectional flow of information.

[0051] Through the multi-layer feature selection and fusion pyramid network processing, the multi-scale feature maps originally from the backbone network are enhanced into fused feature maps rich in high-level semantic information and shallow spatial details, denoted as F3, F4, and F5. At the same time, the shallowest layer P2 feature of the backbone network is also processed by the feature selection module and then fused with the enhanced F3 through the selective feature fusion module to generate the highest resolution enhanced feature map F2.

[0052] In this embodiment, step S4 above utilizes an introduced small target detection head to perform target classification and bounding box regression on the enhanced multi-scale feature map, outputting detection results containing target location information and category information, including: The head network introduces an additional P2 detection head specifically for extremely small targets, building upon the existing detection head. This detection head is connected to the highest-resolution enhanced feature map F2 output from the neck network. The F2 feature map possesses the highest spatial resolution and the richest details in edges, textures, and other aspects. After enhancement fusion through multi-layer feature filtering and a pyramid network, it is rich in both semantic information and spatial details. The P2 detection head directly predicts based on F2 and is specifically designed to detect extremely small targets with pixel sizes smaller than 32×32 pixels.

[0053] Simultaneously, the existing P3, P4, and P5 detection heads perform target classification and bounding box regression on the corresponding scale-enhanced feature maps F3, F4, and F5. Each detection head adopts a decoupled structure, outputting the target category confidence score and bounding box coordinate parameters respectively.

[0054] The outputs of all detection heads are fused and filtered after post-processing operations such as non-maximum suppression to remove redundant detection boxes. The final output includes the target category label and bounding box coordinates, thus completing the accurate identification and localization of small targets in aerial images.

[0055] Example 3: The above is an illustrative scheme of a high-precision aerial small target detection method based on YOLO in this embodiment. It should be noted that the technical solution of a high-precision aerial small target detection system based on YOLO and the technical solution of the above-described high-precision aerial small target detection method based on YOLO belong to the same concept. Details not described in detail in the technical solution of the high-precision aerial small target detection system based on YOLO in this embodiment can be found in the description of the technical solution of the above-described high-precision aerial small target detection method based on YOLO.

[0056] This embodiment also provides a high-precision aerial small target detection system based on YOLO, including: The data preprocessing and architecture building module is used to acquire aerial image data, adjust the images to a uniform size, and build a high-precision aerial small target detection network. The network is based on the YOLO architecture and includes an improved backbone network, an improved neck network, and an improved small target detection head. The multi-scale feature extraction module is used to input aerial image data into a high-precision aerial small target detection network and extract multi-scale feature maps using an improved backbone network. The feature enhancement and fusion module is used to enhance and fuse multi-scale feature maps using an improved neck network to obtain enhanced multi-scale feature maps. The small target detection output module is used to perform target classification and bounding box regression on the enhanced multi-scale feature map using the introduced small target detection head, and outputs detection results containing target location information and category information.

[0057] This embodiment also provides an electronic device applicable to a high-precision aerial small target detection method based on YOLO, including: The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement a high-precision aerial small target detection method based on YOLO, as proposed in the above embodiments.

[0058] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a high-precision aerial small target detection method based on YOLO as proposed in the above embodiments.

[0059] The storage medium proposed in this embodiment belongs to the same inventive concept as the YOLO-based high-precision aerial small target detection method proposed in the above embodiment. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0060] Example 4, refer to Figures 2-5 As an embodiment of the present invention, a high-precision aerial small target detection method based on YOLO is provided. To verify the beneficial effects of the present invention, a simulation experiment is conducted for scientific demonstration.

[0061] Step 1: This embodiment uses YOLO as the baseline model. (Refer to...) Figure 2 The high-precision aerial small target detection model based on YOLO consists of three parts: a backbone network, a neck network, and a head network. The backbone network is used to extract image features, the neck network is used to fuse multi-scale features, and the head network is used for final target classification and localization.

[0062] Step 2: As Figure 3As shown, a multi-scale dilated convolutional residual module is designed. The input feature map of this module has 2C channels. Region residual feature extraction: The input first passes through a 3×3 convolution to expand the number of channels from 2C to 3C, and then passes through BN and ReLU activation functions to obtain the region residual features.

[0063] Multi-scale morphological filtering: The residual features of a region with 3C channels are evenly divided into 3 subgroups along the channel dimension, with each subgroup having C channels. 3×3 depthwise convolutions with dilation rates of 1, 3, and 5 are applied to these 3 subgroups respectively. The outputs of the 3 branches are concatenated along the channel dimension to output a 4C-channel feature map. This map is then compressed back to 2C through a BN layer and a 1×1 pointwise convolution to obtain the semantic residual features.

[0064] Residual connection: The semantic residual features are added element-wise to the initial input features of the DCWR module to obtain the final output of the DCWR module.

[0065] Step 3: As Figure 4 As shown, the MMJD module is constructed. This module uses noisy feature maps. The input is C channels. Sparse Block (SB): Constructs a lightweight 12-layer network, alternating between dilated and standard convolutions, with each layer followed by BN and ReLU. All convolutional kernels are 3×3, maintaining the same number of input and output channels (C). Feature Augmentation Block (FEB): Combines the output of the sparse block with the original input. The feature map is concatenated along the channel dimension to obtain a 2C-channel feature map. It is then passed through a Tanh activation function. Attention Block (AB): The 2C-channel feature map output by the FEB is compressed back to C channels through a 1×1 convolution, generating an attention weight vector. This weight is then element-wise multiplied with the output feature map (C channels) of the FEB to obtain the predicted noise map. Residual denoising: The final sharp features after denoising. Depend on The constructed MMJD module is then inserted into the YOLO backbone network to suppress noise in the feature maps at this stage.

[0066] Step 4: As Figure 5 As shown, an HLFPN is constructed. Taking the P3, P4, and P5 layer features output by the backbone network as an example, their number of channels are C3, C4, and C5, respectively, and the resolution is halved in that order.

[0067] Feature selection module: First, the CA module is applied to P3, P4, and P5 respectively. Taking P4 as an example, global max pooling and global average pooling are performed on it in parallel. The two feature vectors after pooling are added together and fed into the Sigmoid function to obtain the channel weights. Then, they are multiplied with P4 to obtain the filtered P4.

[0068] Then, a 1×1 convolution (FA module) is applied to each of the filtered P3, P4, and P5 to compress their channel number to 256, resulting in N3, N4, and N5.

[0069] Selective Feature Fusion (SFF): Performs top-down fusion. It applies a transposed convolution to N5, upsampling it to the same resolution as N4, to obtain... .Will Input the CA module to generate attention weights Calculate the fused N4 features: Similarly, N4 is upsampled and fused with P3 through the SFF module to obtain N3. Optionally, a bottom-up fusion path can be constructed, downsampling N3 and then fusing it with N4 and P5 again through the SFF module to enhance the bidirectional flow of information. Finally, HLFPN outputs enhanced multi-scale feature maps, denoted as F3, F4, and F5.

[0070] Step 5: Building upon HLFPN, additionally utilize the shallower P2 features from the backbone network. After processing the P2 features by the feature selection module, they are fused with F3 (or directly upsampled from F3) through the SFF module to obtain the enhanced high-resolution feature map F2. In the head network, in addition to the original three detection heads, a P2 detection head specifically for F2 is added. The structure of the P2 detection head is consistent with the decoupled head of YOLO, and it is specifically designed for detecting extremely small targets.

[0071] At this point, the HLAW-YOLOv8 model, including C2f_DCWR, MMJD, HLFPN, and the P2 detection head, has been completed. Through the collaborative work of its modules, this model significantly enhances the detection capability of small targets in aerial images.

[0072] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A high-precision aerial small target detection method based on YOLO, characterized in that, include: Aerial image data is acquired, the images are adjusted to a uniform size, and a high-precision aerial small target detection network is constructed. The network is based on the YOLO architecture and includes an improved backbone network, an improved neck network, and an improved small target detection head. Aerial image data is input into a high-precision aerial small target detection network, and multi-scale feature maps are extracted using an improved backbone network. An improved neck network is used to enhance and fuse multi-scale feature maps, resulting in enhanced multi-scale feature maps. The introduced small target detection head is used to perform target classification and bounding box regression on the enhanced multi-scale feature map, and the output is the detection result containing target location information and category information.

2. The high-precision aerial small target detection method based on YOLO as described in claim 1, characterized in that, The process involves acquiring aerial image data, adjusting the images to a uniform size, and constructing a high-precision aerial small target detection network. This network is based on the YOLO architecture and includes an improved backbone network, an improved neck network, and an improved small target detection head. In the backbone and neck networks, the original C2f modules are replaced by C2f_DCWR modules improved based on multi-scale dilated convolution residual modules; in the backbone network, a multi-module collaborative denoising module is introduced; in the neck network, an improved multi-layer feature screening and fusion pyramid network is used to replace the original feature pyramid structure; and a small target detection head is introduced in the head network. The C2f_DCWR module, which is an improvement based on the multi-scale dilated convolution residual module, extracts multi-scale contextual information by constructing a bottleneck structure and adopting a two-step strategy. The small target detection head constructs the P2 detection head by utilizing the high-resolution feature map of the shallowest layer of the backbone network, based on the multi-layer feature screening and fusion pyramid network.

3. The high-precision aerial small target detection method based on YOLO as described in claim 2, characterized in that, The step of inputting aerial image data into a high-precision aerial small target detection network and extracting multi-scale feature maps using an improved backbone network includes: Aerial image data is input into the backbone network. The input feature map is processed by multiple C2f_DCWR modules in sequence. Each C2f_DCWR module extracts regional residual features and performs standard convolution and activation processing on the input features to generate regional residual features. The regional residual features are divided into multiple subgroups in the channel dimension by multi-scale morphological filtering. Deep convolution with different dilation rates is applied in parallel to each subgroup to obtain multi-scale receptive field information. After feature integration, semantic residual features are formed. The semantic residual features are added element-wise to the initial input, and the processing result is output.

4. The high-precision aerial small target detection method based on YOLO as described in claim 3, characterized in that, The step of inputting aerial image data into a high-precision aerial small target detection network and extracting multi-scale feature maps using an improved backbone network also includes: A multi-module collaborative denoising module is used to suppress and enhance noise in the features. After the noisy input features enter the multi-module collaborative denoising module, they first pass through sparse block processing. The sparse block forms a sparse distribution of features by alternating stacking of dilated convolution and standard convolution. Then, it passes through the feature enhancement block, which concatenates the output of the sparse block with the original input of the module and outputs the fused features after activation processing. Finally, it passes through the attention block, which generates attention weights through convolution operations. These weights are then multiplied element-wise with the output of the feature enhancement block to obtain the predicted noise map. The clear features are then reconstructed through the residual operation between the noisy input features and the noise map.

5. The high-precision aerial small target detection method based on YOLO as described in claim 4, characterized in that, The improved neck network is used to enhance and fuse multi-scale feature maps to obtain enhanced multi-scale feature maps, including: The multi-scale feature maps output from the backbone network are input into the neck network. The neck network adopts a multi-layer feature selection and fusion pyramid structure. The feature selection module and the selective feature fusion module sequentially preprocess and enhance the features of each level. The feature selection module preprocesses each feature map layer. The channel attention submodule in the feature selection module extracts global information of each channel through parallel global pooling operations. After activation, it generates channel attention weights. The channel attention weights are multiplied element-wise with the original features to achieve the selection of key features. The feature alignment submodule in the feature selection module performs convolution operations on each feature map layer to compress the number of channels of each feature map layer to a preset dimension.

6. The high-precision aerial small target detection method based on YOLO as described in claim 5, characterized in that, The method of using an improved neck network to enhance and fuse multi-scale feature maps to obtain enhanced multi-scale feature maps also includes: The selective feature fusion module performs top-down enhancement fusion of features at each level. For any set of high-level and low-level features to be fused, the high-level features are first upsampled to match the spatial resolution with the low-level features. Then, the upsampled high-level features are input into the channel attention submodule to generate semantic attention weights. The semantic attention weights are used to weight and filter the low-level features. Finally, the filtered low-level features are added element-wise to the upsampled high-level features to obtain the fused output feature map.

7. The high-precision aerial small target detection method based on YOLO as described in claim 6, characterized in that, The method utilizes an introduced small target detection head to perform target classification and bounding box regression on the enhanced multi-scale feature map, outputting detection results containing target location and category information, including: The enhanced multi-scale feature map output by the neck network is input into the head network. The head network introduces an additional small target detection head on the basis of the original detection head, which is connected to the highest resolution enhanced feature map output by the neck network. The small target detection head directly predicts based on the highest resolution enhanced feature map. The original detection heads simultaneously perform target classification and bounding box regression on the enhanced feature maps of the corresponding scale. Each detection head adopts a decoupled structure and outputs target category information and bounding box position information independently. The outputs of all detection heads are fused and filtered through post-processing operations to remove redundant detection boxes, and the final output includes the detection results containing the target category label and bounding box coordinates.

8. A high-precision aerial small target detection system based on YOLO, using the method described in any one of claims 1 to 7, characterized in that, include: The data preprocessing and architecture building module is used to acquire aerial image data, adjust the images to a uniform size, and build a high-precision aerial small target detection network. The network is based on the YOLO architecture and includes an improved backbone network, an improved neck network, and an improved small target detection head. The multi-scale feature extraction module is used to input aerial image data into a high-precision aerial small target detection network and extract multi-scale feature maps using an improved backbone network. The feature enhancement and fusion module is used to enhance and fuse multi-scale feature maps using an improved neck network to obtain enhanced multi-scale feature maps. The small target detection output module is used to perform target classification and bounding box regression on the enhanced multi-scale feature map using the introduced small target detection head, and outputs detection results containing target location information and category information.

9. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-executable instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 7.

Images