Fish age reading and otolith analysis using deep learning

Abstract

Fish age reading is a crucial step in the proper management of fisheries. To determine the fish age, several methods have been developed making use of fish structures that give clues on fish growth throughout the different seasons. Supported by years of extensive research and validation efforts, the use of otoliths (ear stones) has become the standard approach. Within otoliths, there are growth rings (annuli) that form patterns through uneven calcium carbonate deposition influenced mainly by seasonal factors affecting the fish growth. Hence, the traditional age reading methodology works by visual inspection and manual counting of these rings to derive the fish age. However, certain cases make the process problematic and error-prone. As these errors have big impact on fisheries management, it is important to explore ways on how these can be prevented. The field of computer vision provides a means to make the process of age reading less reliant on subjective interpretations. Using otolith images, many studies applied classical image processing techniques that take advantage of the changes on image intensity when traversing the otolith from the center to the outer edge. With the progress in artificial intelligence (AI), computer vision methods have taken a new level of sophistication. Using machine learning algorithms, AI models are trained to learn the patterns of growth on the otolith. Early approaches in this direction utilized classical algorithms such as Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs) and some also employed feature engineering in order to create meaningful feature sets (e.g., intensity signals) to be used by the algorithms. Recently, deep learning (DL) algorithms such as Convolutional Neural Networks (CNNs) have gain substantial popularity as they outperform those classical machine learning methods. Early works on DL-based fish age reading have shown excellent accuracy on estimating the fish age based on otolith images. However, the main drawback is that they are formulated as classification or regression making them incompatible with traditional ring counting protocols. In this thesis project, a different perspective for using deep learning on the task of fish age reading was explored. The methods applied were designed to specifically detect and annotate the annuli which are then counted to derive the fish age. Two different object detection and segmentation algorithms were used namely, Mask R-CNN and U-Net. In this thesis, the effectiveness of the two methods was demonstrated along with the tools developed to make the approaches widely accepted by the community. In addition, the study elucidated advanced techniques to improve the accuracy further and also highlighted additional related tasks for general otolith analysis that both algorithms managed to perform effectively.