Marine Remote Sensing Object Detection in Complex Backgrounds Based on Diffusion Models and Multidirectional Spatial Attention

Abstract

Rapid red-tide detection and ship target extraction in nearshore remote sensing images are important for marine ecological monitoring and maritime management. However, nearshore images are easily affected by marine weather conditions and sensor imaging processes, often suffering from cloud cover and stripe noise. These interferences introduce distracting pseudofeatures and weaken target–background discriminability, limiting the stability and robustness of existing detectors. Moreover, the limited number of difficult samples under such interference further increases detection difficulty in complex nearshore scenes. To address these issues, we construct the nearshore complex-scene marine remote sensing dataset (NCSMRS), which contains 1369 images covering three typical nearshore scene types: cloud cover, stripe noise, and clean background, with bounding-box annotations for marine targets. We further propose the diffusion-augmented multidirectional attention detection framework (DMA-Det). DMA-Det adopts layout-guided diffusion augmentation to enrich difficult training samples and introduces a semantic-edge quality filter (SEQF) to improve the reliability of synthetic data through semantic consistency validation and edge quality filtering. In addition, a multidirectional spatial attention (MDSA) module is designed in the detection network to model axial and nonaxial spatial dependencies and recalibrate multiscale features, thereby enhancing target-aware feature representations and suppressing interference from complex backgrounds. Extensive experiments on NCSMRS show that DMA-Det outperforms representative detectors and achieves stable performance across different nearshore scenes, with $\mathrm {mAP}_{50}$ and $\mathrm {mAP}_{50\text {-}95}$ scores of 76.2% and 49.7%, respectively.