A ship detection model with progressive feature fusion and cross-spatial learning attention mechanism for optical remote sensing images

Abstract Using remote sensing images to detect ships is vital for port management, maritime transportation, and security. Challenges such as false detection and omission arise in ship target detection in optical remote sensing images due to the complexity of the background and the diversity of target scales. To address these issues, this paper proposes a novel model called MBE-YOLO. Firstly, the multiscale input spatial pyramid pooling-fast structure is designed to effectively extract more feature information by efficiently integrating the features from different stages of the backbone network. Secondly, the backbone to neck structure is designed with a progressive architecture to mitigate semantic differences between non-adjacent layers in the feature delivery process, thereby significantly reducing the risk of information loss. Finally, we introduce the efficient multi-scale attention attention mechanism, which establishes short and long dependencies through multi-scale parallel subnetworks. This enhances the ability to detect targets in complex environments at various scales. MBE-YOLO is applied to the HRSC2016 and HiresShipDetection datasets. Comparison experiments with current mainstream and state-of-the-art models demonstrate its effectiveness in addressing errors and omissions due to scene complexity and scale variations in remote sensing ship detection, with a parameter size of only 3.24 M.