A Method for Estimating Fish Biomass Based on Underwater Binocular Vision

This paper presents an underwater binocular vision-based fish biomass estimation method using instance segmentation and 3D reconstruction. To validate the system's effectiveness in aquaculture environments and the accuracy of measurement results, three datasets are proposed in experimental and breeding settings. These datasets are respectively used for training and validation of the instance segmentation network; dimension estimation; and mass fitting and estimation. In this work, we introduce pyramid squeeze attention and triple attention in the backbone of instance segmentation network to enhance segmentation accuracy. In the context of free-swimming conditions, where fish often assume non-orthogonal positions relative to the camera's optical axis and exhibit non-linear body shapes, we introduce a stereo disparity map optimization method. This method generates smooth and accurate point cloud and triangular mesh of fish targets, allowing for the estimation of fish fork length and body width based on 3D information. The LAR is also used to estimate the mass by taking the fork length and body width as independent variables.