Automatic fish taxonomy using evolution-constructed features for invasive species removal

Abstract

Assessing the taxonomy of fish is important to manage fish populations, regulate fisheries, and remove the exotic invasive species. Automating this process saves valuable resources of time, money, and manpower. Current methods for automatic fish monitoring rely on a human expert to design features necessary for classifying fish into a taxonomy. This paper describes a method using evolution-constructed (ECO) features to automatically find features that can be used to classify fish species. Rather than relying on human experts to build feature sets to tune their parameters, our method uses simulated evolution to construct series of transforms that convert the input signal of raw pixels of fish images into high-quality features or features that are often overlooked by humans. The effectiveness of ECO features is shown on a dataset of four fish species where using fivefold cross validation an average classification rate of 99.4 % is achieved. Although we use four fish species to prove the feasibility, this method can be easily adapted to new fauna and circumstances.