Coord-Spectral Priors Guided Multisource Feature Fusion Network for Ship Detection in Remote Sensing Images

Detecting ships in complex maritime environments presents significant challenges due to various sources of interference. To address this, we have thoroughly explored complementary information from visible and NIR multi-source sensors, guided by coordinate-spectral priors. The proposed methodology, named VICFT, utilizes coord-spectral prior information to guide multi-source feature fusion, thereby enabling ship detection in scenarios such as cloud and fog occlusion, island interference, sea clutter disturbance, and dense docking environments. The VICFT model consists of three core modules: Visible-NIR Coordinate-Spectral Prior Knowledge Embedding Fusion module VICS. Multi-Scale Cross-Layer Feature Aggregation module MCFA. Multi-Scale SwinTransformer Detection Heads module MSDH. Due to the small size of ship targets in remote sensing images, VICFT uses NWD Loss to represent both the prediction boxes and the ground truth boxes as two-dimensional Gaussian distributions. The Wasserstein distance is employed to measure the similarity between these distributions, which helps to accelerate the convergence of the network. We constructed VIShip, a dataset for detecting ships in visible and NIR under complex scenarios. Finally, a comparative analysis of VICFT with six SOTA methods on the self-constructed dataset VIShip and the publicly available dataset MMShip, respectively, revealed that VICFT achieved the superior precision, recall, mAP50 and mAP50-95 metrics, with enhancements of 0.4%, 0.1%, 1.3%, 5.5% and 0.6%, 0.7%, 0.4%, 6.6%, separately.