An Approach to Automated Estimation of Fish Population Length-Weight Characteristics from Top-View Images in Recirculating Aquaculture Systems

Abstract

This paper addresses the problem of contactless estimation of the length and calculated weight of individual fish from top-view images acquired in recirculating aquaculture system (RAS) tanks. Unlike existing approaches, which are primarily focused on object detection and segmentation, the proposed approach emphasizes the automatic selection of the individual most suitable for subsequent measurement under conditions of high stocking density and mutual fish occlusion. A mathematical image-processing model is proposed in the form of a composition of operators that includes neural-network-based segmentation, geometric analysis of segmented masks, computation of an integral suitability function, estimation of fish length, and calculation of fish weight. Based on the proposed model, a software prototype implementing computer vision and deep learning technologies for automated processing of production video data has been developed. Experimental validation performed on a specialized image dataset of African sharptooth catfish (Clarias gariepinus) confirmed the feasibility of the proposed approach, its capability to automatically select a measurable individual, and the correct operation of the complete software and algorithmic processing pipeline. The obtained results can serve as a basis for the development of intelligent fish monitoring systems and digital twins of aquaculture production processes in recirculating aquaculture systems.