Rethinking Coarse-to-Fine Fully Convolutional Transformers for Salient Maritime Ship Detection in Optical Remote Sensing Imagery

Maritime ship detection in large-scale optical remote sensing imagery is pivotal for rapid anomaly identification, enhancing maritime safety, and promoting sustainable ocean economic development. However, ships in such imagery exhibit extreme multiscale variations, and the imaging mechanism often leads to physical occlusions and indistinct boundaries, which frequently cause high rates of missed detections and false alarms in conventional methodologies. To address these challenges, we introduce the coarse-to-fine saliency-driven maritime ship detection network (C2FSMSDet). This framework first uses a transformer-enhanced coarse detection stage (CoarseDet) to generate robust contextual features and high-fidelity saliency predictions. These predictions subsequently guide a fine-grained instance segmentation stage (FineDet), which performs precise ship delineation by effectively decoupling vessels from confounding boundary elements. Specifically, the CoarseDet stage, built upon a fully convolutional transformer incorporating wide focus block and a criss-cross attention module, extracts high-level semantic features while meticulously considering interpixel correlations to effectively discriminate ships from wake disturbances and background clutter. Subsequently, the FineDet stage, realized through an optimized mask region-based convolutional neural network with a Swin Transformer backbone and a context enhancement module, refines these detections for accurate instance-level results. Comprehensive experimental evaluations on 3 publicly available datasets (airbus ship detection dataset, HRSC2016, and DOTA) demonstrate that the proposed C2FSMSDet significantly outperforms existing state-of-the-art baselines, achieving, for instance, a mean average precision at an intersection over union threshold of 0.5 of 0.953 on a mixed test set.