Feature fusion of body surface and motion-based instance segmentation for high-density fish in industrial aquaculture.

Fish phenotyping serves as a cornerstone for refined management and sustainable development in industrialized aquaculture. High-precision instance segmentation for high-density fish is a key step in obtaining fish phenotypic data; however, it remains challenging in practical production. Given this, an innovative instance segmentation method that synergistically integrates fish body surface features with motion features was developed in this study to tackle the complex issue of high-density fish segmentation in real production effectively. In this method, the Gunnar Farnebäck optical flow is leveraged to express the motion features of fish quantitatively, following which a convolutional neural network of different depths is used to extract the depth feature information of the optical flow images. Simultaneously, the body features output by the backbone network are fused with the motion features. Subsequently, the final fish instance segmentation model was constructed based on SOLOv2. The results of experiments conducted on a self-built dataset of tiger puffer (Takifugu rubripes) revealed that integrating the body surface and motion features resulted in an enhancement of over 12% in the segmentation accuracy of the fish (APm = 48.0%, APl = 84.5%, AP = 84.5%) compared with the baseline (SOLOv2). Compared with other mainstream models of instance segmentation, the AP of the proposed method was improved by 8.3 to 12.8%. Thus, the proposed method can provide technical support for fish phenotypic data in industrialized aquaculture. [ABSTRACT FROM AUTHOR]