Monocular Vision-Based Trout Detection on Floating Cages using YOLOV3

• Published in: 2022 IEEE ANDESCON pp. 1 - 5
• Main Authors: Calisaya, Bryan, Trujillo, Julia, Casas, Luis A. Alfaro, Choquehuayta, Wilder Nina
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.11.2022
• Subjects: Aerospace electronics; Biological system modeling; Computer Vision; Deep learning; Fish; Object Detection; Training; Training data; Trout Detection; Video sequences; YOLO
• DOI: 10.1109/ANDESCON56260.2022.9990003

Trout farming is one of the most important activities, where control and good management are subject to serious errors and waste of time due to the manual way fish farmers carry out this activity. In this research work, a fish detection model is proposed in controlled spaces through video sequence and Deep Learning, with the aim of counteracting the problems that arise in this activity, optimizing the detection process. Accurate detection of trout is the first step to control trout growth estimation, Biomass calculation, and have better organization and management, such as continuous monitoring of fish ponds, amount of feed to be provided to fish, classification and arrangement of fish, etc. In this work, articles referring to fish detection by different authors were reviewed, the proposed model is composed of four main stages which are: 1) data preprocessing where the videos were transformed into images of three frames per second, 2) techniques of data labeling in Make-sense software, 3) data conversion from segmentation to detection, and 4) model training and trout detection. The study results indicate that the proposed model has an accuracy of 72% to detect trout, working with a training data set of 461 images and validation data of 115 images.