An FSFS-Net Method for Occluded and Aggregated Fish Segmentation from Fish School Feeding Images

Abstract

Smart feeding is essential for maximizing resource utilization, enhancing fish growth and welfare, and reducing environmental impact in intensive aquaculture. The image segmentation technique facilitates fish feeding behavior analysis to achieve quantitative decision making in smart feeding. Existing studies have largely focused on single-category object segmentation, ignoring issues like occlusion, overlap, and aggregation amongst individual fish in the fish feeding process. To address the above challenges, this paper presents research on fish school feeding behavior quantification and analysis using a semantic segmentation algorithm. We propose the use of the fish school feeding segmentation method (FSFS-Net), together with the shuffle polarized self-attention (SPSA) and lightweight multi-scale module (LMSM), to achieve two-class pixel-wise classification in fish feeding images. Specifically, the SPSA method proposed is designed to extract long-range dependencies between features in an image. Moreover, the use of LMSM techniques is proposed in order to learn contextual semantic information by expanding the receptive field to extract multi-scale features. The extensive experimental results demonstrate that the proposed method outperforms several state-of-the-art semantic segmentation methods such as U-Net, SegNet, FCN, DeepLab v3 plus, GCN, HRNet-w48, DDRNet, LinkNet, BiSeNet v2, DANet, and CCNet, achieving competitive performance and computational efficiency without data augmentation. It has a 79.62% mIoU score on annotated fish feeding datasets. Finally, a feeding video with 3 min clip is tested, and two index parameters are extracted to analyze the feeding intensity of the fish. Therefore, our proposed method and dataset provide promising opportunities for the urther analysis of fish school feeding behavior.