Date of Award

31-1-2024

Document Type

Thesis

School

School of Electrical & Electroncis Engineering

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

D.Susan

Keywords

Computer Vision, Aquaculture, Deep Learning Networks, Biomass Estimation, Classification Techniques

Abstract

Fish farm is a biosystem with diversity and uncertainties demanding, an intelligent manual monitoring process, which is labor-intensive, invasive, costly, discontinuous, and persistent. The proposed work aims to develop Computer Vision (CV) -based monitoring techniques for precise fish farming to address these problems. Three major monitoring operations namely (i) Fish behavior analysis, (ii) Fish Classification (FC) and (iii) Biomass estimation have been investigated.

Fish schooling activity has been monitored using an overhead vision camera. Problems associated with fish occlusion are addressed using the weighted K-means clustering technique which provides an accurate estimation of fish school locations. Temporal variations of these fish schools are tracked using the Kalman filter (KF)-based multitarget tracking approach. Experimental results illustrate the reliability of the proposed technique to monitor fish school activity in an indoor aquaculture environment.

Multisegmented FC technique using novel fusion-based Deep Learning Network (DLN) architecture is proposed. Inspired by the fish's hydrodynamic nature, convexity deficiency is determined to identify the fish head and is used to segment the fish head, scales, and body. Each segment uses an AlexNet DLN to generate inferences for FC and the inferences are fused using a naive Bayesian fusion layer. Experimental results illustrate a classification accuracy of 98.64% and 98.94% - and Brigham Young University (BYU) datasets respectively. Comparative analysis with other standard networks and ablation studies demonstrates the accuracy and robustness of the proposed fusion architecture, respectively.

DLNs-based segmental analysis technique has been proposed to determine the fish length and convert it to biomass using a calibration curve. In this work, fish segments like head, body, and tail are detected using YOLOv4 (You Look Only Once version-4) DLN. Detected segments are associated using the sequence (head-body-tail) constraint Nearest Neighborhood (NN) technique to define Completely Visible Fish (CVF).

Convex hull and oriented bounding box technique are used to determine the CVF length. Experimental results illustrate a 0.9451 mAP (mean Average Precision) for YOLOv4 and 95.4% of CVFs are detected accurately. Biomass has been estimated with an accuracy of 94.15% and 91.52% for testing and validation image sets, respectively. Integration of the proposed monitoring technique with water quality sensors and feeder systems will be the future extension.

Download

Share

COinS