Automatic Fish Size Estimation from Uncalibrated Fish Market Images Using Computer Vision and Deep Learning

Abstract

Fisheries around the world show an overexploitation, which has led communities to find management strategies to tackle the problem. However, strategies are often taken on the basis of statistical data of dubious real-world utility. To address this problem, accurate biomass extraction calculations are required. The fish market is the place where vessels disembark their catches daily, and therefore a valuable point of contact to retrieve this information. Many small-sized fisheries, as stated by FAO, are a majority in some areas, and small fish markets have more difficulties installing fixed industrial cameras. This paper contributes to these efforts by proposing a complete workflow for fish size regression from uncalibrated images from a mobile camera using fish instance segmentation and classification data provided by a pretrained neural network. Ground truth fish sizes are calculated via homography, and used for comparison. The results show a mean absolute error of 1.7614 ± 2.7633 cm using the CatBoost regressor, and even better at 1.2713 ± 2.0616 cm when considering some calibration parameters at the input.