Deep learning-based segmental analysis of fish for biomass estimation in an occulted environment

Abstract

Fish biomass is one of the reliable parameters that can provide insight into fish and environmental health. Estimation of biomass in a dense and occulted environment is an inevitable and challenging task in modern aquaculture industries, which has been addressed in this work. The proposed work aims to determine the length features of fish using a deep learning-based segmental analysis technique. It tends to analyze the visibility of fish segments like head, body, and tail to define a completely visible fish (CVF). YOLOv4 (You look only once – Version-4) deep learning model is trained and used to detect the fish head, body, and tail segments. The detected segments are associated using sequence constrained nearest neighborhood (NN) association technique guided with fish head orientation. Fish length is estimated using the measurement points identified in the CVF. The measurement point includes head-start, body-center, and tail-end points, which are identified using a convex hull and oriented bounding box (BB). A calibration curve expressing the length-mass relation is used to determine the fish biomass from the estimated length. The proposed methodology is applied to determine the biomass of the genetically improved farmed tilapia (GIFT) fishes in an occulted environment. Experimental results illustrate a 0.9451 mAP of the trained YOLOv4 model and about 95.4% CVFs are detected accurately. A reliable accuracy of 94.15% and 91.52% is observed with testing and validation image sets respectively for biomass estimation.