YOLOv8 algorithm improvement method based on unmanned aerial vehicle aerial image small target detection model

Abstract

The invention belongs to the technical field of computer vision and artificial intelligence, belongs to the cross technical field of target detection, deep learning and image processing, and particularly relates to a YOLOv8 algorithm improvement method based on an unmanned aerial vehicle aerial image small target detection model, which comprises the following steps of: introducing a user-defined feature enhancement module into a YOLOv8 backbone network, a neck part and a detection head part; the self-defined feature enhancement module comprises a context guide self-adaptive fusion module introduced into a backbone network so as to replace part of traditional convolution operation; a space edge sensing feature up-sampling module and a space sensing enhanced convolution module are adopted in the neck fusion network; a fine-grained dynamic pruning detection head is introduced into a detection head detection network. According to the method, the performance of the model in a small target detection scene is effectively enhanced, and the accuracy, robustness and real-time response capability of a detection system are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of target detection, deep learning and image processing, and in particular, it is an improved method of YOLOv8 algorithm based on a small target detection model of UAV aerial images. Background Technology

[0002] In recent years, drone technology has developed rapidly and has been widely used in scenarios such as agricultural monitoring, emergency rescue, security patrol, and environmental assessment. Drones have shown great potential in aerial image acquisition and target detection tasks. However, due to limitations in flight altitude, changes in viewing angle, lighting interference, occlusion, and differences in target scale, the detection of small targets in aerial images still faces significant challenges.

[0003] Existing target detection methods mainly include traditional image processing methods and deep learning-based detection algorithms. Traditional methods rely on manually designed features, have poor robustness, and are difficult to handle complex scenes; while deep learning methods (such as the YOLO series) have high detection performance and are widely used in embedded and real-time systems. YOLOv8, as a representative single-stage detection model, achieves a balance between speed and accuracy and has become one of the mainstream solutions for target detection in UAV images.

[0004] Although YOLOv8 performs well on standard datasets, it still has the following problems in small target detection tasks using drone aerial images:

[0005] 1. Small targets have weak feature representation capabilities and are easily lost during downsampling;

[0006] 2. Multi-scale feature fusion methods are static and struggle to effectively utilize contextual information;

[0007] 3. Insufficient detail restoration capability during the upsampling stage leads to inaccurate positioning;

[0008] 4. Limited receptive field and insufficient ability to model remote context;

[0009] 5. The detection head structure is redundant, resulting in a waste of computing resources; Summary of the Invention

[0010] The purpose of this invention is to solve the problems existing in the prior art and to propose a small target detection algorithm for UAV aerial photography based on an improved YOLOv8. This algorithm achieves efficient detection of small-scale, dense and severely occluded targets by introducing custom feature enhancement modules into the backbone network, neck and detection head of YOLOv8.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] The present invention provides an improved YOLOv8 algorithm based on a small target detection model in UAV aerial images, which introduces a custom feature enhancement module into the backbone network, neck and detection head of YOLOv8;

[0013] Custom feature enhancement modules include:

[0014] The Context-Guided Adaptive Fusion Module (CGAFM) introduced in the backbone network replaces some traditional convolutional operations and is used to adaptively fuse contextual and local information in feature maps.

[0015] The spatial edge-aware feature upsampling module (SEAFUM) and spatial-aware enhancement convolution module (SAEConv) used in the neck fusion network are used to achieve spatial edge-aware upsampling and spatial-aware enhancement of feature maps.

[0016] The fine-grained dynamic pruning detection head (Detect_FGDP) introduced in the detection head network is used for dynamic channel pruning and fine-grained feature enhancement.

[0017] As a further preferred option, the establishment of the Context-Guided Adaptive Fusion Module (CGAFM) includes:

[0018] First, the input feature map is processed by one-dimensional convolution to adjust the channels and map it to the hidden channel space;

[0019] Subsequently, the feature maps are fed into the local context extraction path and the spatial structure perception path, respectively. The former uses standard convolution to capture neighborhood information, while the latter uses dilated convolution to expand the receptive field. The local context extraction path uses standard convolution to capture context information within the neighborhood, while the spatial structure perception path uses dilated convolution to expand the receptive field, thereby enhancing the ability to model the target structural features.

[0020] The outputs of the two paths are weighted and fused, and then normalized and activated to obtain preliminary fusion features.

[0021] This module enables efficient fusion of multi-level features.

[0022] As a further preferred option, to further enhance the model's ability to focus on important channels, fused features are input into the attention modeling module, which includes Simulated Attention Mechanism (SimAM), global average pooling, and two fully connected layers to generate global context information by modeling the dependencies between channels.

[0023] Finally, this contextual information is weighted and combined with the fused features to output a multi-scale feature map with rich semantic expressive power;

[0024] This structural design effectively integrates local and global information, improving the model's detection accuracy and robustness in complex scenarios, and is particularly suitable for drone aerial photography of small targets.

[0025] As a further preferred option, the establishment of the Spatial Edge-Aware Feature Upsampling Module (SEAFUM) includes:

[0026] First, the feature map is input and then compressed to the intermediate channel dimension through a set of depthwise separable convolution operations to reduce computational complexity and extract key semantic information.

[0027] Subsequently, the compressed features are fed into a kernel weight encoder, which predicts the dynamic kernel weights at each upsampling position through lightweight convolution, thereby constructing a reconstructed convolution kernel with content-aware capabilities.

[0028] This module achieves an efficient fusion of feature enhancement and detail preservation.

[0029] As a further preferred solution, in order to enhance the boundary representation capability of small targets, the module introduces an edge-aware branch. This branch uses single-channel convolution to extract edge information from the input feature map and obtains an edge attention map through the Sigmoid function. After upsampling, it is aligned with the main path, that is, upsampling is used to align the output size.

[0030] During the feature reconstruction process, the original feature map is upsampled to the target size through bilinear interpolation, then unfolded into local neighborhood patches through the Unfold operation. The predicted reconstruction kernel weights and the unfolded features are then fused pixel by pixel to complete high-quality upsampling. Finally, the fusion result is weighted and integrated with the edge attention map to achieve a reconstruction output that emphasizes both detail enhancement and edge preservation, i.e., the output feature map has both detail enhancement and edge preservation characteristics.

[0031] Through this structural design, the SEAFUM module can adaptively adjust the upsampling kernel size and fuse multi-scale semantics and edge information, improving the accuracy of small target detection while maintaining the model's lightweightness and efficiency.

[0032] As a further preferred option, the establishment of the Spatial Aware Enhanced Convolutional Module (SAEConv) includes:

[0033] First, a spatial rearrangement operation is performed on the input feature map. By sampling the original input according to different spatial starting positions to form four sub-maps, and then stitching them together in the channel dimension, the spatial information is reorganized and compressed. This operation achieves the reorganization and compression of the spatial information of the feature map without introducing additional computation, thereby enhancing the explicit expression of local structural features.

[0034] Then, the rearranged feature map is input into the Efficient Channel Attention (ECA) module. This module adaptively models the global dependencies between channels through one-dimensional convolution, realizes the weighted emphasis of important semantic features, and improves the attention response of key regions. After obtaining the channel-enhanced feature representation, the module further uses dilated convolution to expand the receptive field while maintaining the spatial resolution, thereby effectively fusing contextual information over a larger range to assist in target structure modeling.

[0035] Finally, a 1×1 convolution layer is used to refine the feature channels, outputting a feature map with stronger expressive power.

[0036] The overall structure, through the process design of "spatial rearrangement + channel modeling + expanded receptive field + channel refinement", effectively improves the model's ability to perceive local details and global structure, and shows better performance in small object detection and complex texture scenes. This module takes into account both local structure modeling and global semantic expression, and improves the feature representation ability in small object detection and complex backgrounds.

[0037] As a further preferred option, the establishment of the fine-grained dynamic pruning detection head (Detect_FGDP) includes:

[0038] By enhancing information flow between features of different scales through dynamic interaction, and introducing FineGrainedLayer to finely model local details, the network can significantly improve the detection accuracy of small targets and targets in complex scenes while maintaining its lightweight nature. This module adopts multi-layer convolution and distributed regression mechanisms in the classification and regression branches, respectively, and combined with DFL decoding and adaptive anchor box generation strategies to achieve more accurate target localization and category discrimination capabilities.

[0039] This module is used to enhance the information interaction between features at different scales and to improve the detection accuracy of small targets and complex scenes by fine-grained modeling local details.

[0040] Beneficial effects: The technical solution of this invention improves the detection accuracy of small targets in UAV aerial images, especially in scenarios with diverse target scales, dense occlusion, and complex backgrounds. It can achieve fast and accurate target localization, and has good practicality and promotional value. Specifically, it has the following advantages:

[0041] 1. This invention addresses the characteristics of densely packed small targets, large scale variations, and susceptibility to occlusion in UAV aerial images, significantly improving detection accuracy and robustness. By introducing a context-guided adaptive fusion module, the feature fusion method can be dynamically adjusted based on contextual information at different levels, alleviating the problems of semantic inconsistency and key feature loss in traditional fusion methods, and effectively enhancing the semantic synergy between features at different scales.

[0042] 2. This invention also designs a spatial edge-aware feature upsampling module. This module enhances the edge details and spatial distribution perception capabilities of the feature map during the upsampling process, improving the model's recognition accuracy for small target boundaries, thus providing better fine-grained perception capabilities in high-resolution scenes. Simultaneously, the proposed spatial perception-enhanced convolution module integrates multi-level contextual information through spatial dimension information rearrangement and expansion of the convolution kernel's receptive field, effectively improving the feature map's response capability to distant and occluded targets.

[0043] 3. To address the information redundancy problem in the detection head stage, this invention introduces a fine-grained dynamic pruning detection head, which can adaptively select and retain key channels based on the feature distribution of the input image. Through joint optimization of channel importance measurement and pruning mechanism, the computational load and parameter quantity are effectively reduced while ensuring detection accuracy, thus accelerating the inference speed of the model. Attached Figure Description

[0044] Figure 1 This is a flowchart of the detection method of the present invention;

[0045] Figure 2 Network architecture diagram of CSSF-YOLOv8 in this invention;

[0046] Figure 3 This is a network structure diagram of the CGA of the YOLOv8 network of the present invention;

[0047] Figure 4 This is a diagram of the SEAFUM network structure of the YOLOv8 network of the present invention;

[0048] Figure 5 This is a diagram of the SAEConv network structure of the YOLOv8 network of the present invention;

[0049] Figure 6 This is a diagram of the FGDP network structure of the YOLOv8 network of the present invention;

[0050] Figure 7 shows the MAP curves before and after the improvement in this invention. Figure 7(a) is the MAP curve of the baseline model YOLOv8, and Figure 7(b) is the MAP curve of the improved model CSSF-YOLOv8. Detailed Implementation

[0051] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0052] This invention proposes an improved YOLOv8 algorithm based on a small target detection model for UAV aerial images, aiming to enhance detection performance in complex scenes captured by UAVs. Addressing the issue of insufficient feature representation capability in traditional YOLOv8 when detecting small, dense, or occluded targets, this invention optimizes the algorithm in four aspects: feature fusion, upsampling reconstruction, receptive field enhancement, and detection head pruning.

[0053] Specifically, a Context-Guided Adaptive Fusion Module (CGAFM) is introduced into the YOLOv8 backbone network to enhance the representation of target context information and local features. In the neck network, a Spatial Edge Aware Feature Upsampling Module (SEAFUM) and a Spatial Aware Enhanced Convolutional Module (SAEConv) are introduced to enhance the preservation of spatial structure information and edge features, thereby improving the feature reconstruction capability of small targets. A fine-grained dynamic pruning detection head (Detect_FGDP) is introduced in the detection head section. Through dynamic channel pruning and attention mechanisms, the detection accuracy of small targets is improved and false detections are reduced. This invention effectively enhances the model's performance in small target detection scenarios and significantly improves the accuracy, robustness, and real-time response capability of the detection system.

[0054] This invention is specifically achieved through the following technical solutions:

[0055] This invention first obtains and preprocesses publicly available drone aerial image datasets, and then constructs a CSSF-YOLO network model based on YOLOv8;

[0056] The detection model introduces a context-aware fusion module (CGA) into the backbone network. The CGA module uses an attention mechanism to enhance edge information in low-level features and adaptively adjusts the fusion weights of features at different levels, effectively reducing redundant information and improving the ability to express key details, thereby enhancing the backbone network's ability to extract small target features at different scales.

[0057] The neck network first introduces a semantic edge-aware feature upsampling module (SEAFUM). This module integrates edge awareness mechanisms and content-aware reorganization strategies during the upsampling process. It captures edge information through a lightweight convolutional structure and completes spatial reconstruction by combining high-level semantic features, thereby enhancing the boundary representation ability of small targets in high-resolution feature maps. Secondly, a spatially aware enhanced convolutional module (SAEConv) is introduced. This module uses a spatial rearrangement strategy to break the local perception limitation of traditional convolutions, guiding the network to focus on a wider range of contextual information. It improves the abstraction and representation ability of features without significantly increasing the computational cost, effectively enhancing the recognition performance of small targets in complex backgrounds.

[0058] A fine-grained dynamic pruning (FDPH) detection head module is introduced into the head network. This module dynamically filters based on the importance of feature channels, eliminating redundant channels and improving inference speed while enhancing the detection head's adaptability to multi-scale targets. The collaborative design of these modules enables the constructed detection model to possess stronger feature extraction, fusion, and representation capabilities, significantly improving the detection accuracy and robustness of small targets in complex aerial photography scenes. Finally, the CSSF-YOLO model is trained and validated using training and validation sets. The test set is then input into the finally trained detection model to obtain detection results and evaluate model performance.

[0059] This invention relates to a method for detecting small targets in drone aerial images, and the specific steps are as follows:

[0060] S1: As Figure 1 We obtained the publicly available VisDrone 2019 dataset and cleaned, transformed, and normalized it. The preprocessed dataset was divided into training, validation, and test sets, containing 6471, 548, and 1610 images respectively, serving as data for object detection from the drone's perspective. This data was used to train the YOLO series object detection networks, facilitating model training and evaluation.

[0061] S2: As Figure 2 A small target detection model for UAV aerial photography was constructed based on the YOLOv8 algorithm. Its network structure diagram is shown below. Figure 2 As shown, the UAV aerial small target detection model, namely the improved CSSF-YOLO algorithm, includes a backbone feature extraction network, a neck fusion network, and a head detection network.

[0062] S3: As Figure 3 By introducing a context-guided adaptive fusion module (CGAFM) into the backbone of the original YOLOv8 network model, an improved YOLOv8 network model is obtained, which adaptively fuses contextual and local information in the feature maps.

[0063] The specific steps for using the CGAFM module are as follows:

[0064] S31: The input feature map is first downsampled by a 3×3 convolution with a stride of 2, reducing the spatial size to half of the original size and increasing the number of channels by 2 times (2×C), thus completing the initial feature compression and scaling transformation.

[0065] S32: The downsampled feature map is fed into two branches simultaneously: the local context extraction branch obtains local detailed features within the neighborhood through depthwise separable convolution; the global context modeling branch expands the receptive field with dilated convolution to effectively capture a wider range of contextual information.

[0066] S33: The feature maps extracted from the two paths are fused by element-wise weighting. The fusion result is then processed by batch normalization and activation function to enhance the stability and nonlinearity of the feature representation.

[0067] S34: The fused features are further refined through 1×1 convolutions to distribute the channels, and then fed into the lightweight channel attention module FGlo to model the dependencies between channels and improve the ability to focus on key semantics.

[0068] S35: Finally, the attention-enhanced features are added element-wise to the output of the residual branch to achieve information preservation and enhancement, and output a downsampled feature map with stronger semantic and structural expressive capabilities.

[0069] Design a Context-Guided Adaptive Fusion Module (CGAFM) to replace some traditional convolutional operations for efficient fusion of multi-level features in object detection networks. The module first receives the input feature map and maps it to the hidden channel space through a set of one-dimensional convolutional operations. Subsequently, the feature map is fed into a local context extraction path and a spatial structure awareness path, respectively. The local path uses standard convolutions to capture contextual information within the neighborhood, while the spatial path uses dilated convolutions to expand the receptive field, thereby enhancing the ability to model the structural features of the target. The outputs of the two paths undergo weighted fusion and normalized activation processing to obtain preliminary fused features.

[0070] To further enhance the model's ability to focus on important channels, fused features are input into the attention modeling module. This module includes a simulated attention mechanism (SimAM), global average pooling, and two fully connected layers, generating global contextual information by modeling the dependencies between channels. Finally, this contextual information is weighted and combined with the fused features to output a multi-scale feature map with rich semantic expressiveness. This structural design effectively integrates local and global information, improving the model's detection accuracy and robustness in complex scenes, and is particularly suitable for small target detection tasks in UAV aerial photography.

[0071] like Figure 3As shown, the CGA module preserves key contextual information during the fusion of high-level and low-level features. Traditional fusion methods such as Concat or weighted summation can easily lead to the burying of key details during information transmission. The CGA module introduces an attention mechanism to enhance edge information in low-level feature maps and adaptively adjusts the fusion weights of features at different levels, thereby reducing feature redundancy and improving information utilization efficiency.

[0072] S4: As Figure 4 By introducing a spatial edge-aware feature upsampling module (SEAFUM) into the neck network of the original YOLOv8 network model, an improved YOLOv8 network model is obtained.

[0073] The specific steps for using the SEAFUM module are as follows:

[0074] S41: The input feature map is first compressed using depthwise separable convolution. A combination of 3×3 depthwise convolution and 1×1 pointwise convolution is used to extract key features and reduce computational cost, generating an intermediate feature representation.

[0075] S42: The compressed intermediate features are fed into a convolutional encoder to predict the content-aware reassembly weights corresponding to each pixel position. The kernel size is k_enc, and the number of output channels is related to the kernel size and the upsampling ratio. Then, pixel shuffle is used to restore it to the upsampling space.

[0076] S43: To preserve detailed information, the edge enhancement branch convolves the input feature map and generates an edge attention map, which is then interpolated to the upsampling size for subsequent weighted enhancement.

[0077] S44: The input feature map is upsampled to the target size using bilinear interpolation, while an Unfold operation is used to extract local neighborhood feature blocks in preparation for reconstruction. This process adaptively adjusts the kernel size according to the current image size, balancing accuracy and efficiency.

[0078] S45: The local features obtained by Unfold are weighted and fused with content-aware weights. The einsum operation is used to implement the pixel-level dynamic convolution process to obtain the fused feature map.

[0079] S46: Finally, multiply the fused result by (1 + edge attention map) to enhance the importance of image edge regions, improve the response to small or edge targets, and output the final upsampled feature map.

[0080] This module reconstructs spatial details of the upsampled feature map by enhancing edge information and content awareness, effectively improving the ability to perceive spatial details of small targets.

[0081] like Figure 4 As shown, the SEAFUM module is specifically designed for the upsampling stage, introducing an edge-aware mechanism and a content-aware reorganization strategy. In the process of recovering high-resolution image details from low-resolution feature maps, the SEAFUM module captures edge location information through a lightweight convolutional structure and adaptively reorganizes spatial features in conjunction with semantic content, effectively improving the accuracy and completeness of small target feature representation.

[0082] S5: As Figure 5 By introducing a spatially aware augmented convolutional module (SAEConv) into the neck network of the original YOLOv8 network model, an improved YOLOv8 network model is obtained.

[0083] The specific steps for using the SAEConv module are as follows:

[0084] S51: Spatial downsampling segmentation. The input feature map is divided into four sub-sampled feature maps according to spatial location (even row even column, odd row even column, even row odd column, odd row odd column), and concatenated in the channel dimension to complete spatial downsampling and increase the channel dimension.

[0085] S52: Channel attention enhancement. The spliced ​​feature map dynamically adjusts the channel weights through the ECA attention mechanism to enhance the expressive power of key feature channels.

[0086] S53: Dilated convolution expands the receptive field. Dilated convolution is used to perform convolution operations on attention-weighted features to expand the receptive field and effectively capture a wider range of contextual information.

[0087] S54: Channel refinement and adjustment. Features are compressed and adjusted through 1×1 convolution to reduce redundancy and improve the compactness of feature representation.

[0088] like Figure 5 As shown, the SAEConv module is placed at the neck connection of the backbone network to expand the receptive field and improve feature abstraction capabilities. This module breaks the local perception limitation of convolution through a spatial rearrangement strategy, guiding the network to focus on a wider range of contextual information without significantly increasing computational cost, thereby enhancing the ability to distinguish small targets in complex backgrounds.

[0089] S6: As Figure 6 A fine-grained dynamic pruning detection head module (FDPH) is introduced to process multi-scale feature maps. The channel importance of the input multi-scale features is used to dynamically select effective channels, perform dynamic channel pruning and fine-grained feature enhancement, and obtain the detection head of the head network of the improved YOLOv8 network model.

[0090] The specific steps for using the FDPH module are as follows:

[0091] S61: Channel compression is performed on the input multi-layer feature map through a 1×1 convolutional layer to reduce the computational complexity of the model and adapt to the fine-grained feature requirements of small targets in UAV aerial images.

[0092] S62: The DyHead module is introduced to perform dynamic feature fusion and enhancement of multi-layer features, thereby improving the model's ability to represent features of targets at different scales.

[0093] S63: Combines a fine-grained layering module, refines spatial information through multiple sets of convolution operations, and enhances the expression of texture and edge details of small targets.

[0094] S64: Employs a multi-layer convolutional structure (cv2 and cv3) to extract richer contextual information and category discrimination features, while integrating the DFL (Distribution Focal Loss) mechanism to improve boundary regression and classification accuracy, achieving lightweight and high-precision detection.

[0095] like Figure 6 As shown, the FDPH module, a key improvement to the detection head, incorporates a fine-grained pruning mechanism. This mechanism dynamically filters channels based on their importance in feature maps at different scales, effectively eliminating redundant channels, improving inference speed, and enhancing the detection head's adaptability to multi-scale targets through feature selection.

[0096] S7: The detection model composed of the improved YOLOv8 network model obtained in steps S3 to S6, the backbone network of the improved YOLOv8 network model, the head network of the improved YOLOv8 network model, and the neck network of the improved YOLOv8 network model together constitutes the improved YOLOv8 algorithm method based on the small target detection model of UAV aerial image.

[0097] The improved YOLOv8 network model was trained using a publicly available drone aerial image dataset as the training set. A comparative experiment was conducted between the original and improved YOLOv8 network models. The experiment involved training both models on a self-made lung nodule dataset for 150 iterations. The operating environment was: CUDA 11.6; Python 3.8.18; GPU NVIDIA GeForce RTX 3090.

[0098] Output the corresponding detection results and evaluate them:

[0099] To verify the effectiveness of each improved module in the detection system of this invention, the original YOLOv8l-P2 network model was used as a benchmark, with precision, recall, mean precision, and parameter count as evaluation metrics. Ablation experiments were designed using different combinations of multiple improved modules. The experimental results are shown in Table 1.

[0100] Table 1 Comparison of Ablation Experiment Results

[0101]

[0102] To comprehensively evaluate the specific contributions of each proposed improved module to target detection performance, this paper designed multiple ablation experiments based on the YOLOv8l backbone network. The experimental results are shown in Table 1. First, the improved spatial residual convolution module (SREConv) was integrated into YOLOv8l-P2 to form model YOLOv8l-P2_1. Compared with the original model, its AP50 increased from 0.483 to 0.505, and AP increased from 0.296 to 0.310, indicating that SREConv has a significant effect on enhancing local feature representation and suppressing redundant information. Subsequently, the lightweight upsampling module LECARAFE was added to obtain YOLOv8l_2, with AP50 and AP slightly improved to 0.508 and 0.312, respectively. This shows that LECARAFE can improve the ability to preserve fine-grained information during upsampling and further optimize the feature fusion effect.

[0103] Based on YOlOv8l_2, the FGDP module was introduced to construct YOlOv8l_3, which improved the AP50 to 0.524 and AP to 0.321, a significant improvement over the previous version. This verifies the effectiveness of FGDP in modeling global context and dynamic receptive field, and helps improve the model's robustness to multi-scale targets. Finally, based on YOlOv8l_3, a context-guided attention mechanism (CGA) was further integrated to construct the final model YOlOv8l_ours, achieving the highest detection accuracy with an AP50 of 0.529 and AP of 0.326, an improvement of 4.6 points compared to the base model. This fully demonstrates the role of CGA in enhancing spatial attention and improving the detection accuracy of small targets.

[0104] Overall, each module positively impacts model performance in different ways, with SREConv and FGDP providing particularly significant improvements. The combination of multiple modules exhibits good complementarity and synergistic enhancement, enabling the final ensemble model to significantly improve detection performance while maintaining low computational overhead, validating the effectiveness and practicality of the proposed improvement strategy. Experiments on the VisDrone2019 dataset demonstrate that the proposed improved YOLOv8 algorithm exhibits stronger robustness in UAV aerial images with dense small targets, severe occlusion, and varying scales. Compared to the baseline model YOLOv8l-P2, the improved algorithm achieves a 4.6% improvement in detection accuracy AP50 and a 3% improvement in AP, validating the effectiveness and sophistication of the proposed method in complex real-world application scenarios.

[0105] To more fully evaluate the improved YOLOv8 algorithm method for small target detection based on UAV aerial images in this invention, this detection method was compared with several other models under the same configuration and dataset, as well as the basic YOLOv8l-P2 network model. The experimental results are shown in Table 2:

[0106] Table 2 Comparative Experiments

[0107]

[0108] As shown in Table 2, to verify the effectiveness of the proposed improved YOLOv8-L-P2 target detection method in UAV image target detection tasks, the applicant constructed a systematic comparative experiment, evaluating the performance of the proposed method against several existing mainstream detection algorithms on a standard UAV image detection dataset. The comparison methods cover four main categories: traditional detectors for UAV images, end-to-end detectors, real-time end-to-end detectors, and general real-time detectors. The relevant results are as follows: Figure 2 As shown in the table.

[0109] In traditional UAV detector categories (such as QueryDet, ClusDet, DCFL, and HIC-YOLOv5), although some methods (such as ClusDet) achieve an AP50 score of 50.6%, their overall detection accuracy (AP) remains insufficient. For example, QueryDet's AP is 28.3%, and HIC-YOLOv5's is only 26.0%. In contrast, as shown in Table 2, the method of this invention, YOLOv8-L-P2(ours), achieves an AP score of 32.6% and an AP50 score of 52.9%, significantly outperforming these methods and demonstrating the higher accuracy and robustness of the model in handling small-scale, high-density UAV target detection tasks.

[0110] In end-to-end detector comparisons, while mainstream DETR and RT-DETR series methods possess certain detection performance (e.g., RT-DETR-R50 has an AP of 28.4%), their model parameters are generally large, and their computational complexity is high, making them unsuitable for embedded deployment. In contrast, the method of this invention maintains a high AP (32.6%) and AP50 (52.9%) while keeping computational resource consumption within an acceptable range, exhibiting superior deployment adaptability and inference efficiency.

[0111] Among real-time end-to-end detectors for UAV imagery, while the UAV-DETR-R50 model performs well in terms of detection accuracy (AP 31.5%, AP50 51.1%), it is still slightly lower than the performance metrics achieved by the method of this invention. Although the method of this invention has a relatively high number of parameters while maintaining leading detection accuracy, it is practically feasible to adapt to current mainstream AI hardware platforms.

[0112] Further comparison with current mainstream YOLO series real-time detection methods (such as YOLOv8, YOLOv9, YOLOv10, etc.) reveals that the YOLOv8-L-P2(ours) proposed in this invention achieves significant improvements in all detection accuracy metrics. For example, compared to the original YOLOv8-L-P2 model, AP increases from 29.6% to 32.6%, and AP50 increases from 48.3% to 52.9%, an average improvement of 4.6 percentage points, demonstrating the clear advantages of this invention's method in target detection tasks.

[0113] The introduction of the above modules resulted in an improvement of AP50 and AP of 4.6% and 3% respectively on the VisDrone2019 dataset, demonstrating stronger scene adaptability and practical application value.

[0114] In summary, this invention, through improvements to the YOLOv8 model structure, achieves higher detection accuracy, stronger adaptability, and acceptable computational overhead in UAV image target detection tasks, demonstrating significant technical effects and practical application value.

[0115] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An improved YOLOv8 algorithm method based on a small target detection model in UAV aerial images, characterized in that: Custom feature enhancement modules are introduced into the backbone, neck, and detection head of YOLOv8; Custom feature enhancement modules include: The Context-Guided Adaptive Fusion Module (CGAFM) introduced in the backbone network replaces some traditional convolutional operations and is used to adaptively fuse contextual and local information in feature maps. The spatial edge-aware feature upsampling module (SEAFUM) and spatial-aware enhancement convolution module (SAEConv) used in the neck fusion network are used to achieve spatial edge-aware upsampling and spatial-aware enhancement of feature maps. The fine-grained dynamic pruning detection head (Detect_FGDP) introduced in the detection head network is used for dynamic channel pruning and fine-grained feature enhancement.

2. The improved YOLOv8 algorithm method based on a small target detection model in UAV aerial images according to claim 1, characterized in that: The establishment of the Context-Guided Adaptive Fusion Module (CGAFM) includes: First, the input feature map is processed by one-dimensional convolution to adjust the channels and map it to the hidden channel space; Subsequently, the feature maps are fed into the local context extraction path and the spatial structure perception path, respectively; the local context extraction path uses standard convolution to capture contextual information within the neighborhood, while the spatial structure perception path uses dilated convolution to expand the receptive field. The outputs of the two paths are weighted and fused, and then normalized and activated to obtain preliminary fusion features.

3. The improved YOLOv8 algorithm method based on a small target detection model in UAV aerial images according to claim 2, characterized in that: To further enhance the model's ability to focus on important channels, fused features are input into the attention modeling module, which includes Simulated Attention (SimAM), global average pooling, and two fully connected layers to generate global contextual information by modeling the dependencies between channels. The contextual information is weighted and combined with the fused features to output a multi-scale feature map with rich semantic expressive power.

4. The improved YOLOv8 algorithm method based on a small target detection model in UAV aerial images according to claim 1, characterized in that: The establishment of the Spatial Edge Aware Feature Upsampling Module (SEAFUM) includes: First, the feature map is input and then compressed to the intermediate channel dimension through a set of depthwise separable convolution operations to reduce computational complexity and extract key semantic information. Subsequently, the compressed features are fed into a kernel weight encoder, which predicts the dynamic kernel weights at each upsampling position through lightweight convolution, thereby constructing a reconstructed convolution kernel with content-aware capabilities.

5. The improved YOLOv8 algorithm method based on a small target detection model in UAV aerial images according to claim 4, characterized in that: To enhance the boundary representation capability of small targets, an edge-aware branch is introduced in the Spatial Edge-Aware Feature Upsampling Module (SEAFUM). This branch extracts single-channel edge information from the input feature map through a set of convolution operations and obtains an edge attention map via the Sigmoid function. The edge attention map is then upsampled to the target resolution through bilinear interpolation to achieve spatial alignment with the output of the main branch.

6. The improved YOLOv8 algorithm method based on a small target detection model in UAV aerial images according to claim 5, characterized in that: To achieve high-quality restoration of the main features and synergistic enhancement of edge information, the module further performs feature reorganization operations in the main branch: In the feature reorganization process, the input feature map is first upsampled to the target size using bilinear interpolation. Then, the upsampled feature map is unfolded into local neighborhood patches through a dynamically initialized Unfold operation. Combined with the reorganization convolution kernel weights predicted by the content-aware encoder, a pixel-wise weighted fusion operation is performed to generate a preliminary high-resolution feature map.

7. The improved YOLOv8 algorithm method based on a small target detection model in UAV aerial images according to claim 1, characterized in that: The establishment of the Spatial Aware Enhanced Convolutional Module (SAEConv) includes: First, the input feature map is spatially rearranged by sampling the original input at different spatial starting positions to form four sub-maps, which are then concatenated along the channel dimension to reorganize and compress spatial information. Then, the rearranged feature map is input into the efficient channel attention module (ECA). This module adaptively models the global dependencies between channels through one-dimensional convolution to achieve weighted emphasis on important semantic features and improve the attention response of key regions. After obtaining the channel attention-enhanced feature map, the module further employs dilated convolution to expand the receptive field while maintaining the spatial resolution. Finally, a 1×1 convolution layer is used to refine the feature channels, outputting a feature map with stronger expressive power.

8. The improved YOLOv8 algorithm method based on a small target detection model in UAV aerial images according to claim 1, characterized in that: The establishment of the fine-grained dynamic pruning detection head (Detect_FGDP) includes: By enhancing information flow between features at different scales through dynamic interaction, and introducing FineGrainedLayer to finely model local details, the network can significantly improve the detection accuracy of small targets and targets in complex scenes while maintaining its lightweight nature.

9. The improved YOLOv8 algorithm method based on a small target detection model in UAV aerial images according to claim 8, characterized in that: This module employs multi-layer convolution and distributed regression mechanisms in the classification and regression branches, respectively, in conjunction with DFL decoding, to achieve more accurate target localization and category discrimination capabilities.

10. The improved YOLOv8 algorithm method based on a small target detection model in UAV aerial images according to claim 9, characterized in that: By combining anchor box generation strategies, the detection accuracy and robustness of multi-scale targets in complex scenes can be further improved.

Images