An algorithm for detecting dense small objects in aerial photography based on coordinate position attention module

• DOI: 10.1049/ipr2.13061
• Publication Date: 2024-02-21T08:47:01Z
• Authors: Huixin Wu; Yang Zhu; Mengdi Cao
• Keywords: Position (finance); Backbone network; Feature (linguistics); Aerial image

Abstract To address the challenges of detecting a large number objects and high proportion small in aerial drone imagery, an dense object detection algorithm called coordinate position attention module you only look once (CPAM‐YOLO) is proposed based on (CPAM). In backbone network CPAM‐YOLO, CPAM embedded that decomposes channel into two 1D feature encoding processes, selectively combines features each through weighted sum all features. Finally, are aggregated along spatial directions, increasing effective information utilization input positions channels. The network, enhancement heads have been optimized to improve accuracy while ensuring lightweight network. Using networks significantly reduce parameters using high‐resolution retain more semantic detailed algorithm's performance was evaluated publicly available VisDrone2019 dataset. Compared baseline YOLOv5l, CPAM‐YOLO achieved 4.5% improvement mAP 0.5 3.2% 0.95 . These experimental results demonstrate strong practicality for image.