Deep Neural Networks for Otolith Identification

Abstract

Otoliths identification is the main metrics for assessment and management of fish stocks. Its analysis forms an essential part of the routine work by fisheries scientists. This task requires precision and accuracy through representing a significant cost to fisheries management. In the last two decades, many attempts to use computer vision to otolith identification have been carried out. Several classification features and machine learning algorithms have been used; however, up-to-date, none has the computer assisted identification systems have out-performed human experts. Furthermore, most of these methods suffer from over-learning problems. This paper aims to train deep neural networks (DNN) to identify otolith using a convolutional neural network with a maximum accuracy and minimum loss while taking into account the problem of overfitting. The proposed CNN architecture has been trained and tested on an augmented database of otoliths, which contains 15405 images belonging to 15 classes. Our experimental results suggest that our designed CNN has a good accuracy in training and vulnerable to overfitting. The evaluation result of the proposed method shows that the performance of our method is close to that of biologists .It achieves 98.5% of test accuracy with minimum loss accuracy 0.08.