Fish Morphological Feature Recognition Based on Deep Learning Techniques

Abstract

The object features in an image or a video frame can be determined as a number of landmarks detected with the assistance of Deep Learning techniques. In this paper, object detection and image segmentation are initially performed to isolate a fish in an image and then eight landmarks are aligned in order to measure the fish dimensions and the position of its mouth and fins. Four popular Mediterranean fish species have been used in this study: Merluccius merluccius (cod fish), Dicentrarchus labrax (sea bass), Sparus aurata (sea bream) and Diplodus puntazzo. The first three of these species are grown in fish farms. For this reason, monitoring the morphological features of these fishes in their environment is of particular interest for ichthyologists and the proposed method can serve this purpose. The proposed method has been developed using Convolution Neural Networks and OpenCV in Python and MATLAB applications. The accuracy in the estimation of the fish dimensions for an initial data set with 20 images/species, ranges between 80% and 91% while the fins are located with an accuracy of up to 93%.