A lightweight unmanned aerial vehicle small target detection method based on DCC-YOLOv11

Abstract

This invention discloses a lightweight UAV target detection method based on DCC-YOLO, belonging to the field of UAV small target detection. This detection method primarily addresses the challenges of small targets, complex backgrounds, severe occlusion, and limited computational resources in UAV aerial images. The invention enhances high-level feature extraction capabilities, expands the receptive field, and generates feature maps with richer feature information by introducing an improved dilated convolution C3k2 module. Secondly, a context-guided module is introduced to suppress background noise and enhance the model's perception of target regions. Finally, a lightweight cross-channel feature modulation mechanism is introduced in the feature fusion stage to improve cross-channel information interaction and nonlinear expression capabilities. When applied to UAV small target detection, this improved method exhibits advantages over the original YOLOv11n method, including higher accuracy, fewer parameters, lower computational cost, and faster computation speed, making it highly competitive.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target detection technology, and in particular to a method for small target detection in drone aerial images, which belongs to the application of lightweight deep learning models in intelligent recognition of drone images. Background Technology

[0002] With the widespread application of drones in fields such as inspection, agricultural monitoring, and disaster assessment, the vision systems they carry place higher demands on the accurate detection and recognition of targets in images. However, due to limitations in drone shooting altitude, viewing angle, and system computing resources, targets in aerial images often exhibit characteristics such as small target size, complex background, severe occlusion, and limited resources, making traditional target detection algorithms prone to missed detections and false detections in such scenarios.

[0003] The YOLO series of algorithms are widely used in target detection tasks due to their end-to-end and efficient detection capabilities, especially lightweight versions such as YOLOv11n, which have already achieved a certain level of edge deployment capability. However, their performance in small target detection in complex backgrounds still has room for improvement, particularly in areas such as limited receptive field, weak context awareness, and insufficient utilization of cross-channel information. Therefore, there is an urgent need to design a novel small target detection method that balances detection accuracy and model lightweightness to meet the complex requirements of UAV image scenarios.

[0004] Furthermore, existing methods often suffer from unstable and insufficient feature extraction when faced with complex interference factors in UAV aerial images, such as varying target scales, similar foreground and background features, and limited image quality (e.g., blurring, illumination variations). This severely impacts the accuracy and robustness of detection results. Especially when small targets occupy few pixels, have blurred boundaries with the background, and lack contextual information, conventional convolutional structures struggle to effectively perceive their position and semantic information. Some studies have attempted to improve model perception capabilities by increasing network depth or introducing attention mechanisms, but this inevitably leads to increased computational complexity and deployment difficulties. For resource-constrained edge computing scenarios (such as embedded UAV systems), highly complex models struggle to meet real-time performance and power consumption control requirements. Therefore, achieving efficient perception, accurate localization, and reliable identification of small targets under limited computing power has become a critical problem that urgently needs to be solved in UAV target detection. Summary of the Invention

[0005] The purpose of this invention is to provide a lightweight UAV small target detection method based on DCC-YOLOv11. It addresses the problems of small target size, weak features, severe occlusion, and limited computing resources by improving the YOLOv11n network structure, thereby improving detection accuracy, enhancing feature perception capabilities, and maintaining the model's high efficiency and lightweight characteristics.

[0006] To achieve the above objectives, the present invention proposes the following technical solution:

[0007] Step S1: Deploy the object detection environment for model training, ensuring that the hardware platform meets the minimum configuration requirements;

[0008] Step S2: Obtain a publicly available drone image dataset containing various scenes and target categories, and divide the dataset into two parts, with the majority of the data used as the training set and the minority of the data used as the test set;

[0009] Step S3: Design and construct the network structure of the DCC-YOLO target detection model. The network structure includes an image preprocessing module, a lightweight backbone feature extraction network, a feature fusion neck network, and a multi-scale detection head module to achieve effective extraction and detection of multi-scale small targets in UAV images.

[0010] Step S4: Based on the training set of the dataset obtained in step S2, train the DCC-YOLO model and optimize the model parameters through multiple rounds of iteration to obtain a small target detection model with better detection accuracy.

[0011] Step S5: Use the test set in the dataset to evaluate the performance of the trained DCC-YOLO optimal model, comprehensively examine its detection accuracy, recall and inference speed, verify the practicality of the model in the small target detection scenario of UAV images, output the detection results and realize the localization and classification of targets in the image.

[0012] The DCC-YOLO detection model established in step S3 is as follows:

[0013] Step S3.1: Preprocess the original image dataset by using strategies such as Mosaic stitching enhancement, MixUp blending enhancement, image scaling and translation to stitch multiple images into one, increasing the diversity of the training data.

[0014] Step S3.2: The backbone network adopts an improved version of the CSP-Darknet53 network, which extracts feature information of the input image through Conv module convolution to form a feature map;

[0015] Step S3.3: Replace the original C3k2 module in YOLOv11 with the dilated convolution module C3k2_DWR. The module integrates the DWR (Dilation-wise Residual) structure, expands the receptive field by combining dilated convolution, and alleviates the problem of deep cross-scale information loss and gradient vanishing by using residual connections, thereby enhancing the ability of high-level semantic features to model small targets.

[0016] Step S3.4: Introduce the Contextual Guidance Block (CGBlock) to replace part of the traditional Conv convolutional layer. Through the context-aware mechanism, improve the model's attention to key regions, realize the dynamic modeling and separation of target region and background region features, and optimize the accuracy of target information extraction.

[0017] Step S3.5: Aggregate multi-scale features using the SPPF (Spatial Pyramid Pooling - Fast) module, and then feed the fused features extracted by the backbone network into the neck network.

[0018] Step S3.6: By using the C2PSA module, the pyramid slice attention mechanism (PSA) is integrated to dynamically adjust the channel feature response, strengthen the focus on key regions, and improve the detection accuracy of small targets and the multi-scale feature fusion effect.

[0019] Step S3.7: In the fusion path of the Neck structure, a lightweight cross-scale feature fusion module (CCFM) is introduced. Combining 1×1 convolution and dynamic weighting mechanism, it enhances feature representation ability while reducing redundant parameters and computational overhead, thereby improving detection accuracy while taking into account computational efficiency.

[0020] Step S3.8: Extract the filtered information using the C3k2_DWR module and pass it to the head network.

[0021] Step S3.9: The detection head adopts the native decoupled detection head structure of YOLOv11, and the classification and regression tasks are modeled independently. In the classification branch, depthwise separable convolution (DWConv) is used to compress the amount of computation while maintaining the prediction accuracy.

[0022] Step S3.10: Use the non-maximum suppression (NMS) algorithm to quickly remove the excessive redundant bounding boxes obtained in the previous step, and finally obtain the detection results.

[0023] The performance evaluation for step S5 is as follows:

[0024] Step S5.1: First, verify the effectiveness of each improved module through ablation experiments on a public dataset;

[0025] Step S5.2: By conducting comparative experiments with existing mainstream target detection algorithms, the advantages of the proposed method in terms of detection accuracy and computational efficiency are verified;

[0026] Step S5.3: Visualize and analyze the model performance through experimental results data tables and visualized test results.

[0027] In small target detection tasks using drone aerial images, precision, recall, and mean average precision (mAP) are typically used as core evaluation metrics, and their calculation formulas are as follows:

[0028]

[0029]

[0030]

[0031]

[0032] Where Precision represents accuracy, Recall represents recall, n represents the total number of target categories, TP (TruePositive) represents the number of correctly detected targets, FP (False Positive) represents the number of non-targets that are falsely detected as targets, FN (False Negative) represents the number of targets that the model fails to correctly identify, mAP reflects the balance between the model's detection precision and recall across all categories (unit: %), mAP50 refers to the mAP value when IoU is 0.5, mAP@0.5 represents the average detection precision across all categories when the IoU threshold is 0.5, and mAP@0.5:0.95 calculates the average precision at different IoU thresholds from 0.5 to 0.95.

[0033] Besides detection accuracy, the computational complexity and storage requirements of a model are also key factors affecting its practicality. Commonly used computational metrics include the number of parameters and Giga Floating Point Operations per Second (GFLOPs). The number of parameters refers to the sum of the weight parameters of each layer of the model, reflecting the model's storage requirements; GFLOPs represent the number of floating-point operations required by the model during inference, measuring its computational complexity. Lower parameter counts and GFLOPs help reduce computational resource consumption and improve inference speed, which is especially suitable for application scenarios with limited computational resources, such as drones.

[0034] The present invention has the following beneficial effects and advantages:

[0035] (1) By integrating the DWR structure with dilated convolution and residual connection, the receptive field and feature preservation capabilities of the model are significantly improved, the high-level semantic modeling of small targets is enhanced, the gradient vanishing and cross-scale information loss problems in deep networks are effectively alleviated, and the detection accuracy is improved without significantly increasing the complexity of the model.

[0036] (2) The CGBlock module is added to the model to realize context-aware modeling, accurately separate the target area from the background noise, and at the same time, the SPPF and C2PSA modules are used to fuse local and global information and dynamically adjust the channel feature response to enhance the multi-scale fusion quality and significantly improve the recognition accuracy of small targets.

[0037] (3) A lightweight cross-scale feature fusion module, CCFM, is introduced into Neck. It integrates the idea of ​​cross-channel feature modulation and optimizes cross-scale information interaction through a dynamic weighting mechanism, while further controlling the model parameters. The design concept of first lightweighting the model to ensure that the computational load of model parameters is not increased is adopted to improve accuracy as much as possible. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the steps of a lightweight UAV target detection method based on DCC-YOLO according to the present invention.

[0039] Figure 2 This is a flowchart illustrating the implementation of the DCC-YOLO target detection model of the present invention.

[0040] Figure 3 This is a structural diagram of the C3k2 module described in this invention.

[0041] Figure 4 This is a schematic diagram of the DWR module described in this invention.

[0042] Figure 5 This is a structural diagram of the dilated convolution module C3k2_DWR described in this invention.

[0043] Figure 6 This is a structural diagram of the CGBlock context bootstrapping module described in this invention.

[0044] Figure 7 This is a structural diagram of the CCFM module described in this invention.

[0045] Figure 8 This is a comparison chart of the YOLOv11n and DCC-YOLOv11PR curves described in this invention.

[0046] Figure 9 This is the DCC-YOLO visualization detection result described in this invention. Detailed Implementation

[0047] like Figure 1As shown, the technical solution of the present invention includes five steps: environment deployment, data acquisition, model building, model training, and performance testing.

[0048] Step S1 Environment Deployment: Ensure that the hardware platform meets the minimum configuration requirements. The system must have Python 3.8 or above installed, and the versions of PyTorch, CUDA, and graphics cards must be consistent.

[0049] Step S2 involves obtaining the dataset: This study uses the publicly available VisDrone2019 dataset as the source of experimental data. This dataset was collected by the AISKYEYE Laboratory at Tianjin University and contains 8599 still images taken from the perspective of a high-altitude drone, including 6471 images in the training set, 1580 images in the test set, and 548 images in the validation set. It covers various complex urban environments and multiple target categories, demonstrating good representativeness and challenge.

[0050] Step S3 involves model construction: The DCC-YOLO target detection model of this invention is constructed in four parts: data preprocessing, backbone network, neck network, and head. Through collaborative optimization between modules, the model's feature extraction capability and detection accuracy in small target detection tasks are effectively improved.

[0051] Step S4, training the model, involves configuring the following training parameters: input image resolution of 640×640, epochs of 200, batch size of 16, workers of 8, lr0 of 0.01, lrf of 0.01, momentum of 0.937, and weight decay coefficient of 0.0005.

[0052] Step S5 Performance Testing: The test dataset is used to perform small target detection through the model to obtain the detection performance.

[0053] like Figure 2 As shown, the flowchart of the DCC-YOLO target detection model of the present invention is as follows: The present invention introduces an improved dilated convolution C3k2 module (C3k2_DWR) to enhance high-level feature extraction capabilities, expand the receptive field, and generate feature maps with richer feature information; secondly, it introduces a context-guided module (CGBlock) to suppress background noise and enhance the model's perception of the target region; finally, in the feature fusion stage, it introduces a lightweight cross-channel feature modulation mechanism (CCFM) to enhance feature expression capabilities through cross-channel information interaction, and effectively extract key features of small targets while reducing the number of parameters and computational overhead, thereby improving detection accuracy.

[0054] The specific implementation process is as follows:

[0055] Step S1 Environment Deployment: In this embodiment, the object detection environment for model training is first deployed. Specifically, a hardware platform including CPU, GPU, memory, and storage devices is constructed, and an operating system and deep learning framework are installed on the hardware platform. At the same time, the Python runtime environment and related dependent libraries are installed and their versions are matched to ensure the stability and computational efficiency of the model training process.

[0056] Step S2 Data Preprocessing: The images captured by the drone are preprocessed, including operations such as translation, scaling, mixup and mosaic data augmentation on the images in the training data after partitioning, to ensure that the input images meet the input requirements of the YOLOv11 model.

[0057] Step S3 Backbone Network: Extract features from the input image to provide basic data for subsequent detection and classification tasks. The backbone network described in this invention has the following improvements:

[0058] Step S3.1: Replace the Bottleneck module in the C3k2 module with the DWR module to introduce dilated convolution technology to expand the receptive field and capture more contextual information. This is especially helpful in stabilizing the feature extraction process of the network and reducing the impact of external factors on object detection when dealing with varying lighting and complex backgrounds. The C3k2 structure is as follows: Figure 3 As shown, the C3k2 module includes convolutional layers, feature branch structures, feature processing units, and feature fusion units. First, the input feature map is fed into the convolutional layer for feature extraction, resulting in an intermediate feature map. Then, the intermediate feature map is divided into two sub-feature maps along the channel dimension. One sub-feature map is directly transmitted as the main branch; the other sub-feature map is input into the feature processing unit as an auxiliary branch for feature extraction. When the parameter C3k is False, the feature processing unit consists of several Bottleneck structures connected in series; when the parameter C3k is True, the feature processing unit consists of several C3k structures connected in series, with the number of concatenations being n. After feature processing, the outputs of the main branch and the auxiliary branch are concatenated along the channel dimension, and channel fusion and feature reshaping are completed through a convolutional layer to obtain the final feature map.

[0059] Step S3.2: As Figure 4As illustrated in the diagram, the DWR module employs a residual connection structure. First, the input feature map is fed into a 3 × 3 convolutional layer and processed by batch normalization (BN) and ReLU activation to obtain an initial feature map. Then, the initial feature map is divided into three sub-feature maps along the channel dimension, with each sub-map containing C / 2 channels, and these are fed into three parallel branches. Each branch contains a 3 × 3 dilated depth convolutional layer with dilation rates set to d=1, d=3, and d=5, respectively, to extract feature information within different receptive fields. After channel concatenation of the feature maps output from each branch, they are fed into a batch normalization layer for normalization, followed by channel fusion through a 1 × 1 pointwise convolutional layer to obtain a fused feature map. Finally, the fused feature map and the input feature map are added element-wise through a residual connection to obtain the output feature map, thus achieving effective fusion of multi-scale features while preserving the original feature information.

[0060] Step S3.3: As Figure 5 As shown, based on the C3k2 module structure, the following is introduced: Figure 4 The DWR module shown is improved to construct the C3k2_DWR structure. Specifically, after the input feature map is processed by a convolutional layer, it is divided into two sub-feature maps along the channel dimension. One sub-map is directly transmitted as the main branch, while the other sub-map is used as an auxiliary branch for feature extraction through multiple DWR modules. After feature processing, the feature map output from the auxiliary branch is concatenated with the feature map from the main branch along the channel dimension and then input into subsequent convolutional layers for channel fusion and feature reshaping to obtain the output feature map. Through this structure, the effective extraction of multi-scale feature information is achieved while controlling the complexity of network layers, thereby enhancing the high-level feature representation capability.

[0061] Step S3.4: As Figure 6As shown, a newly proposed context-guided module, CGBlock, is used to optimize the limitations of traditional convolutional operations in YOLOv11 and improve the network's adaptability to complex scenes by introducing a context-guided mechanism. This module employs a multi-scale information fusion strategy to enhance the integration of local and global features, especially in small object detection, improving the capture of global information and reducing false negatives and false negatives. CGBlock consists of several key components that work together to integrate local, surrounding, and global contextual information to improve feature representation. Local Feature Extractor (floc(*)): Extracts features from local regions (e.g., the red region (a)). Surrounding Context Extractor (fsur(*)): Extracts features from surrounding regions (e.g., the yellow region (b)), providing relevant contextual information. Joint Feature Extractor (fjoi(*)): Combines local features and surrounding contextual features to generate joint features, fusing local and contextual information. Global Context Extractor (fglo(*)): Extracts global contextual features, combining information from the entire image or a wider region (e.g., the green region (c)). The final output is combined through element-wise multiplication, where local features, surrounding context features, and global context features work together to generate a stronger feature representation, which helps improve the detection and classification capabilities of the target region.

[0062] Step S3.5: As Figure 7 As shown, based on the characteristics of small targets in UAV-view images, this paper proposes an improved feature fusion network based on a Cross-Scale Feature Fusion Module (CCFM) to enhance feature representation capabilities, improve detection accuracy, and maintain computational efficiency. The network framework mainly consists of three parts: a dynamic channel adaptive mechanism, a lightweight fusion module, and a multi-level interaction strategy. A 1×1 convolutional layer is introduced into the feature fusion path to dynamically adjust the channel distribution of the input feature map, reducing redundant parameters through dimensionality reduction, while simultaneously enhancing the feature representation capability for dense small targets through cross-channel information interaction. Subsequently, a lightweight fusion module composed of N RepBlocks is used, replacing traditional convolutions with structural reparameterization technology to reduce computational complexity while preserving detailed information of multi-scale features. In the cross-level interaction stage, deep semantic features are upsampled and concatenated with shallow high-resolution features through a concatenation operation, preserving local details of small targets such as pedestrians and vehicles; simultaneously, an element-wise add strategy is combined to fuse the global contextual semantics of multi-scale features, mitigating feature loss caused by target occlusion. Finally, the feature dimensions are unified and output through the Flatten operation.

[0063] Step S3.6: The detection head adopts a decoupled detection head structure, with classification and regression tasks modeled independently. In the classification branch, depthwise separable convolution (DWConv) is used to compress the computational load while maintaining prediction accuracy. At the same time, the non-maximum suppression (NMS) algorithm is used to quickly remove the excessive redundant bounding boxes obtained from the previous step, and finally the detection results are obtained.

[0064] Step S4: Train the model: Train the constructed DCC-YOLO model. Set the training parameters as follows: input image resolution 640×640, number of training epochs 200, batch size 16, number of data loading threads (workers) 8, initial learning rate (lr0) 0.01, learning rate decay factor (lrf) 0.01, momentum parameter 0.937, and weight decay factor 0.0005. Iteratively train the model under the above parameter configuration, and select the optimal model weights based on the validation set performance to obtain the optimal model for UAV small target detection.

[0065] Step S5 Performance Testing: Perform performance testing on the trained model. Input the test dataset into the optimal model obtained through training to perform object detection, obtain the model's prediction results on the test set, and calculate performance indicators such as detection accuracy, recall rate, and inference speed based on the prediction results to evaluate the model's detection performance in the UAV small target detection task.

[0066] like Figure 8 As shown, a comparison chart of the DCC-YOLO and YOLOv11n PR curves is presented.

[0067] like Figure 9 As shown, the visualization detection results of DCC-YOLO and YOLOv11n are demonstrated.

[0068] In summary, the lightweight UAV target detection method based on DCC-YOLO proposed in this invention demonstrates superior target recognition capabilities in complex scenes compared to traditional models. Addressing the potential issues of missed and false detections of small targets in the YOLOv11n model when processing UAV aerial images, this invention significantly improves the model's receptive field coverage and feature modeling capabilities for small targets through multiple structural optimizations, strengthening its focus on key regions. While maintaining low computational cost, this model can more effectively handle multi-scale information fusion and spatial modeling tasks, achieving precise feature selection and enhanced representation, thereby comprehensively improving the accuracy and robustness of small target detection.

[0069] The above-described embodiments of the present invention are merely one specific implementation method. Any features or steps not explicitly excluded can be equivalently replaced or adjusted according to actual needs. Unless there is a logical conflict or technical mutual exclusion, all functional modules, structural features and method steps disclosed in this specification can be combined and reconstructed in any form.

Claims

1. A lightweight UAV target detection method based on DCC-YOLO, characterized in that... The steps include: Step S1: Deploy the model training environment; Step S2: Obtain a publicly available dataset of small target images from multiple scenarios and categories of UAVs. Divide the dataset into two parts, with the majority of the data used as the training set and the minority as the test set. Step S3: Design the DCC-YOLO network architecture and construct a lightweight DCC-YOLO UAV target detection model. Step S4: Iteratively train the model using the training set and the DCC-YOLO network architecture to optimize the model parameters and obtain the optimal small target detection model. Step S5: Evaluate the performance of the optimal model using the test set to ensure that the model meets the predetermined accuracy and real-time processing requirements. Finally, output the small target detection results to obtain the localization and classification of targets in the UAV images.

2. The lightweight UAV target detection method based on DCC-YOLO as described in claim 1, characterized in that, The DCC-YOLO network architecture includes a preprocessing section, a backbone network, a neck network, and a detection head section.

3. The lightweight UAV target detection method based on DCC-YOLO as described in claim 2, characterized in that, The preprocessing section performs preprocessing operations on the original image dataset, using Mosaic mosaic enhancement, mixup enhancement, image scaling and translation to stitch multiple images into one, increasing the diversity of the training data.

4. The lightweight UAV target detection method based on DCC-YOLO as described in claim 2, characterized in that, The backbone network uses the DWR module to replace the original C3k2 module in YOLOv11 to design a brand new C3k2_DWR module. A context-guided module is introduced to replace some traditional convolutional layers (Conv). Through the context-aware mechanism, the model’s attention to key regions is improved, and the dynamic modeling and separation of target region and background region features are realized, thereby optimizing the accuracy of target information extraction.

5. The lightweight UAV target detection method based on DCC-YOLO as described in claim 2, characterized in that, The neck network employs a cross-scale feature fusion module to enhance feature representation capabilities while controlling the number of parameters to improve overall lightweighting.

6. The lightweight UAV target detection method based on DCC-YOLO as described in claim 2, characterized in that, The detection head section employs a decoupled head structure and uses depthwise separable convolution to compress computational load while maintaining prediction accuracy.

7. The lightweight UAV target detection method based on DCC-YOLO as described in claim 1, characterized in that, In the performance evaluation of small target detection tasks in UAV aerial imagery, accuracy, recall, and mean precision are typically used as core evaluation metrics, and their calculation formulas are as follows: Where Precision represents accuracy, Recall represents recall, n represents the total number of target categories, TP represents the number of correctly detected targets, FP represents the number of non-targets that were falsely detected as targets, FN represents the number of targets that the model failed to correctly identify, and mAP reflects the balance between the model's detection precision and recall across all categories, in units of:

Images