Non-contact weight estimation system for fish based on instance segmentation

Non-contact methods for estimating cultured fish weight are essential for aquaculture companies to develop aquaculture strategies and management plans. However, it is challenging for current technologies to precisely measure the size of fish in a densely breeding environment and estimate the weight of fish because of issues including occlusion, fish bending and not facing the camera. This study, which focuses on the aforementioned issues, suggests a technique for creating an instance segmentation dataset appropriate for measuring fish size as well as an attention-based fully convolutional instance segmentation network (CAM-Decoupled-SOLO) to extract fish contour features. This study uses a concatenation approach to fuse low-level features and high-level features in order to safeguard low-level features so as to address the issue of low-level feature information loss in the bottom-up propagation process of feature pyramid networks. Additionally, by combining pixel location infor-mation with channel attention mechanism, this work achieves the integration of target location information and channel dimension information. To validate the developed methodology, it has been applied to an aquaculture system. The results showed that the mAP of proposed method would be improved by 2.6% compared with Decoupled-SOLO. Both qualitative and quantitative assessments would be enhanced when compared to the existing well-liked one-stage network. Moreover, to save labor cost and avoid fish damage, an automatic fish perimeter measurement model and a weight prediction system were constructed by combining fish contours extracted by CAM-Decoupled-SOLO and binocular vision. There was a significant correlation between fish perimeter and weight, which can be used to estimate the weight of fish in complex aquaculture environments.

Automated summary

Non-contact methods for estimating cultured fish weight are important for aquaculture companies, but current technologies face challenges in precisely measuring fish size and estimating fish weight in dense breeding environments. This study proposes a technique for creating a dataset suitable for measuring fish size and an attention-based network for extracting fish contour features. The study addresses the issue of low-level feature loss by combining features from different levels and integrates target location information and channel dimension information. The developed methodology has been successfully applied to an aquaculture system and shows improved performance compared to existing approaches.