Fully automatic system for fish biomass estimation based on deep neural network

Abstract

The approach for estimating biomass in non-contact, free-swimming fish has encountered difficulties such as fish body occlusion, bending, non-orthogonal angles, and low efficiency. To address these issues, this study had combined fish posture recognition (using deep learning technology) with biomass estimation (utilizing stereo vision technology) for the first time, and developed a fast, precise, and fully automatic fish biomass estimation system. The improved single-stage target detection algorithm significantly improved the direct detection and extraction of high-quality images of moving fish in real-time, eliminating the need for manual processing of images that may have imperfect posture. Fish body length and height were measured in the real world using binocular stereo vision technology. Finally, the fish body weight can be estimated by considering the relationship among their body length, height, and weight. The experiment confirmed that the system had successfully avoided influencing factors that could affect fully automatic estimation. The results demonstrated a strong linear relationship between the estimated and measured fish body weights, with a mean relative error (MRE) of 2.87%. There were no significant differences between the estimated and measured weights (p = 0.94). The MRE of the multi-factor model was much lower than that of the single-factor model (length-weight of 8.86% and height-weight of 7.41%). The results indicate that the system developed is a highly effective approach to fully automated biomass estimation. This can be used to guide actual production and further study of the mechanism of fish growth.