Automatic fish counting from underwater video images: performance estimation and evaluation

Abstract

Cabled observatories offer new opportunities to monitor species abundances at frequencies and durations never attained before. When nodes bear cameras, these may be transformed into the first sensor capable of quantifying biological activities at individual, populational, species, and community levels, if automation image processing can be sufficiently implemented. Here, we developed a binary classifier for the fish automated recognition based on Genetic Programming tested on the images provided by OBSEA EMSO testing site platform located at 20 m of depth off Vilanova i la Gertrú (Spain). The performance evaluation of the automatic classifier resulted in a 78% of accuracy compared with the manual counting. Considering the huge dimension of data provided by cabled observatories and the difficulty of manual processing, we consider this result highly promising also in view of future implementation of the methodology to increase the accuracy.