CKDR-YOLO: A Convolutional Kernel Dilated Residual-Based Detector for Visible-Light Ship Detection

Abstract

To address insufficient small-target perception and poor environmental adaptability in UAV-based visible-light maritime ship detection, we propose convolutional kernel dilated residual YOLO (CKDR-YOLO), an efficient real-time detector built upon YOLOv11. A convolutional kernel dilated residual (CKDR) module integrating adaptive convolution, multiscale dilated convolutions, and residual feature fusion is designed to enhance local feature extraction and global contextual perception. To verify practical deployment capability, the proposed method was further validated on a laboratory-developed programmable UAV platform for real-time maritime monitoring. Experimental results show that CKDR-YOLO improves precision, recall, mAP@50, and mAP@50–95 by 0.3%, 5.6%, 3.6%, and 3.6% over YOLOv11, respectively, while achieving real-time inference at 101 frames/s. The proposed framework demonstrates robust performance under challenging maritime conditions, indicating its effectiveness for airborne intelligent ship detection.