M-FSDistill: A Feature Map Knowledge Distillation Algorithm for SAR Ship Detection

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 17; pp. 13217 - 13231
• Main Authors: Wang, Guohui, Qin, Rui, Xia, Ying
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Background noise; Biological system modeling; Computational modeling; Datasets; Decoupling; Deep learning; Detectors; Distillation; Distilling; Feature extraction; Feature imitation; Feature maps; Image enhancement; Image resolution; knowledge distillation; Knowledge engineering; Knowledge representation; Machine learning; Marine vehicles; Modules; multiscale fusion; Radar detection; SAR (radar); ship detection; Synthetic aperture radar; synthetic aperture radar (SAR) image
• ISSN: 1939-1404; 2151-1535; 2151-1535
• DOI: 10.1109/JSTARS.2024.3426288

Limited by the capacity and computing ability of platform payload, researchers often face the challenge of balancing lightness and performance of models in synthetic aperture radar (SAR) ship detection, especially for models based on deep learning. Nonetheless, traditional lightweight methods, such as reduced convolutional layers and pruning, can easily lead to missed detections in models. Researchers have introduced knowledge distillation algorithms to address the issue of poor performance of lightweight models. However, the improvement effect of algorithms is limited due to shortcomings such as noise interference in the background and improper distillation strategies, especially for small ship detection with complex backgrounds. Aiming to address the limited performance improvement of distillation algorithms and missing detections of small ships in distillation models, we propose a multiscale feature enhancement and foreground-scene feature distillation algorithm for SAR ship detection. Specifically, in order to improve distillation efficiency, the feature learning distillation module is proposed to improve the quality of distillation knowledge by separating foreground and scene distillation. Then, the ship feature representation enhancement module utilizes a feature map decoupling and attention-based multiscale fusion algorithm to enhance student model's learning of small ship features and reduce missing detection. To validate the performance of the proposed method, we conducted experiments on SAR ship detection dataset (SSDD) and high-resolution SAR images dataset (HRSID) datasets and compared with several advanced methods. The results indicate that the models using our algorithm achieved significant improvements in average precision (AP). For instance, on the SSDD dataset, RetinaNet, Cascade R-CNN, and RepPoints based on ResNet18 achieved AP scores of 95.5%, 95.4%, and 95.9% respectively, surpassing the baseline by 3.9%, 3.1%, and 1.9%.