Efficient semi-supervised learning model for limited otolith data using generative adversarial networks

Otolith shape recognition is one of the relevant tool to ensure the sustainability of maritime resources. It is used to study taxonomy, age estimation and discrimination of stocks of fish species. The most performant otolith image classification models are based on convolutional neural network approaches. To build an efficient system, these models require a large number of labeled images, which is hard to obtain. The lack of data became a big challenge, and a real problem of otolith images classification models, it causes the over-fitting issue, which is the main trouble of deep convolutional neural network based models. In this paper, we present a relevant solution for the insufficiency of data. We propose a new semi-supervised classification model based on generative adversarial network. Our results showed that the model is more efficient and also perform better than convolutional neural network system even with a small training dataset. With this efficiency and performance, we found in addition that the accuracy of the model reached 80% on training set of say, 75 images compared to other models such as a convolutional neural network model which accuracy is limited to 60%.