Ship detection in haze and low-light remote sensing images via colour balance and DCNN

Abstract

Ship detection in marine remote sensing images (RSI) is an area of growing interest with a wide range of potential applications. However, current ship detection methods often suffer from low accuracy and poor robustness when dealing with RSI due to the influence of diffuse haze or low-light conditions. To overcome these issues, this study develops a hybrid deep learning framework for detecting ships in RSI, consisting of three main components: RSI dehaze, RSI enhancement, and an improved YOLO-v5s network. Firstly, the RSI dehaze uses colour space transformation to correct the colour cast of blurred RSI and performs dehaze using transmission mapping and haze-line clustering. Then the weakly illuminated RSI is enhanced by utilising illumination mapping between input and enhanced images to describe the correlation between images. Finally, the ship detection model is constructed by improving the YOLO-v5s model with channel split and shuffle, which effectively suppresses redundant features and accurately detects ships in RSI. The HRSC2016 and DIOR datasets are used for comprehensive performance assessment. The experimental results demonstrate that the proposed framework achieves superior ship detection precision, with mAP, Pre and Rec reaching 95.27%, 97.55% and 99.74%, respectively, on the HRSC2016 dataset.