C4D-YOLOv8: Improved YOLOv8 for Object Detection on Drone-captured Images

Unmanned aerial vehicles (UAVs), as an emerging technology with vast application prospects, produce distinctive images due to factors, such as their shooting environment, altitude, and equipment. These UAV-captured images possess unique characteristics compared with traditional pictures, including high resolution, abundant presence of small objects, and object density. However, most existing object detection networks primarily focus on detecting large objects, resulting in limitations in terms of detecting small targets. The current detection networks frequently lose significant information whilst extracting features of small targets, resulting in a substantial decrease in detection accuracy. Accordingly, we have employed an improved version of YOLOv8 to address these issues, taking into consideration the aforementioned challenges,. We have devised a novel feature extraction module called C4, which aims to regain the crucial feature information of small targets that may have been overshadowed by the background. In contrast to the previous C2F or C3 modules, the C4 module incorporates a pathway with a larger convolutional kernel. Additionally, we have introduced a small target detection layer and an attention mechanism-based target detection head called ‘DyHead’ into the original yolov8s architecture. This approach has yielded a remarkable improvement in performance, achieving a mean average precision value of 45.8% on the VisDrone dataset, which represents a 4.9% accuracy enhancement compared with YOLOv8s (baseline). Subject classification codes: include these here if the journal requires them