Publication

Abstract : A Multiscale Attention and Refinement Integrated Ship-Detection Network (MARIS-Net) has been proposed as a hybrid framework intended for accurate ship detection in Synthetic Aperture Radar (SAR) imagery. The architecture has been derived from the U-Net family but redesigned with a multi-scale attention mechanism so that both fine ship boundaries and larger maritime structures are captured in a balanced manner. To increase localization accuracy, particularly in detecting small or adjacent vessels, a contour-refinement module based on OpenCV morphological operations has been included. Extensive data augmentation was applied during training so that the model became resistant to common SAR distortions such as speckle noise, backscatter fluctuation, and sea clutter. The evaluation was carried out on two standard benchmarks, namely the High-Resolution SAR Images Dataset (HRSID) and the SAR Ship Detection Dataset (SSDD). On these datasets, precision of 95.77 % and 99.47 %, recall of 95.87 % and 97.69 %, and mean Average Precision (mAP) of 89.13 % and 99.15 %, respectively, have been obtained. The main contribution of this work lies in combining multi-scale attention-based segmentation with contour-oriented refinement, a strategy that closes the gap between conventional U-Net structures and newer attention-based designs. Overall, MARIS-Net has demonstrated reliable performance for maritime surveillance, defense, and remote-sensing operations, even under weather and visibility conditions where optical sensors provide limited information.