Use of Computer Vision Related Technologies to Identify Fish and their Anomalies within a Group

Abstract

Conservation of the aquatic species and their habitats has become important than ever and the authorities and laws alone are not enough to achieve a success. The lack of knowledge among the general public about the importance and the value of aquatic species compared to the interest shown by them on the conservation of land animals like elephants, is considerably low. Another barrier in terms of conservation is the lack of updated information on species. Hence the focus of the research is to bridge the knowledge gap and develop a feasible mechanism for better data collection by using computer vision related technologies. The solution has three components, a mobile app, a node API and an analytics dashboard backed by a cloud db. Since the lack of knowledge on species was identified a mobile app was developed to classify fish without internet. After comparing SVM methods and deep learning methods to develop a classifier it was evident that the deep learning will be providing better accuracy. The GoogleNet and MobileNet were shortlisted as both were depth wise separable CNNs which has a lesser computational cost than the traditional CNNs. After training both CNNs and comparing model size vs accuracy MobileNet with 1.0 𝛼 and 224 𝜌 was chosen as the best suited classifier for the task which is delivering about 70% accuracy. The app also features a fish log tool which is designed to gather data with a community driven approach and the dashboard represent the analytics for the collected data. To bridge the knowledge on the unclassified species a mechanism to connect experts in the fish domain and users has been developed. Hence the research provides an overall solution to the two identified barriers in aquatic conservation.