Global Recurrent Mask R-CNN: Marine ship instance segmentation

Abstract

In intelligent ship navigation, instance segmentation technology is considered an accurate and efficient tool for vision perception in marine scenarios. However, the complex sea surface background and the diversity of ship types and sizes in marine environments pose significant challenges for instance segmentation, especially for small-scale targets. Therefore, this paper presents an end-to-end Global Recurrent Mask R-CNN (GR R-CNN) algorithm designed to enhance the multi-scale segmentation performance of ship instances in marine settings. Initially, this method proposes the Recurrent Enhanced Feature Pyramid Network (RE-FPN) module, which uses a feature recurrence and bidirectional chaining fusion mechanism to deeply integrate both deep and shallow features of images, effectively extracting multi-scale features and semantic information. Subsequently, we propose the Fine-Grained Global Fusion Mask Head (FGFMH) module, utilizing a fine-grained multi-layer receptive field extraction mechanism to enhance the extraction of global and multi-scale features. These two modules collaborate to further improve the ship instance segmentation capability. Experiments conducted on the MS COCO test-dev, PASCAL VOC, and custom OVSD datasets demonstrate accuracy improvements of 1.8%, 3.29%, and 1.3%, respectively, compared to Mask R-CNN. Our method surpasses various advanced techniques and provides valuable insights for the research on multi-scale instance segmentation of ships in complex environments.