TGC-YOLO: Detection Model for Small Objects in UAV Image Scene

In the field of drone target detection, existing object detection models often have problems of missed detection and false detection when dealing with challenges such as scale changes, dense targets and irregular arrangement of small targets. Therefore, this paper proposes a new detection model TGC-YOLO. Firstly, the EMA attention mechanism module is introduced to encode global information, and pixel-level features are further aggregated through dimensional interaction, which effectively reduces the attention to irrelevant information and maintains good feature extraction capabilities. Secondly, in terms of network structure, the FasterBlock module is used to replace the traditional BottleNeck module in C2f in YOLOv8n, so as to design a lightweight FasterC2f module, which reduces the number of parameters while ensuring the accuracy of the model. Thirdly, the Slim-neck structure is adopted on the Neck layer, and the model's detection ability for small objects is enhanced through the combination of GSConv and VoV-GSCSP. Then, the TGC three-channel feature fusion structure is designed based on GSConv to effectively aggregate multi-scale feature maps, enhance the feature fusion ability of the model, and improve the performance of small target detection. Finally, the dynamic head (DyHead) based on the attention mechanism is introduced to further improve the accuracy of small target detection by combining the three attention mechanisms of scale, space and task. Compared with YOLOv8n, TGC-YOLO has achieved 7.7%, 6.8% and 7.4% improvement in mAp@50, Precision and Recall respectively. Meanwhile, the model size is reduced by 0.1MB, enabling fast and accurate detection of small targets.