Xingyu Wang∗ and Wei Huang∗
Small target detection, YOLO, feature extraction, attentionmechanism, cross-scale feature fusion∗ School of Computer Science and Engineering, Wuhan Institute ofTechnology, Wuhan 430205, China; e-mail: [email protected];[email protected] author: Wei HuangRecommended by Gian Luca Foresti, Ph.D.
Drone aerial imagery presents a particularly challenging task for small object detection in the field of computer vision, primarily due to the extremely small size of targets, their often blurry appearance, and dense distribution. These challenges lead to significant issues of missed detections and false alarms in existing models, while their high computational demands restrict practical deployment on mobile platforms. To address these challenges, this paper proposes a novel KLC-YOLO algorithm. The core contributions of this study include the design of a C2f-KAN module that replaces standard convolutions with KAN convolutions, enabling more efficient parameter utilisation and significantly enhancing fundamental feature extraction capabilities. Additionally, a large-Kernel separable attention mechanism is introduced in the SPPF layer to expand the receptive field and focus on critical features, thus improving multi-scale small object perception. An efficient cross-scale feature fusion module is also developed to leverage shallow high-resolution feature information while reducing model complexity. Furthermore, a P2 small object detection layer is added, specifically designed to capture the smallest targets in images. Comprehensive experiments on the authoritative VisDrone2019-DET dataset demonstrate that KLC-YOLO significantly outperforms baseline models in both accuracy and efficiency. Specifically, the proposed model achieves 39.9% mAP50 and 24.1% mAP50:95 with only 2.55-MB parameters, representing a 13% reduction compared to YOLOv8n, and achieving absolute performance gains of 6.6% and 4.7%, respectively. These results validate the effectiveness and superiority of KLC-YOLO in achieving more accurate and lightweight small object detection for drone applications in complex scenarios.
Important Links:
Go Back
