Crosspooled FishNet: transfer learning based fish species classification model

Abstract

Fish species classifiction is an important task for biologists and marine ecologists to frequently estimate the relative abundance of fish species in their natural habitats and monitor changes in their populations. Traditional methods used for fish species classifiction were laboriuos, time consuming and expensive. So, there is need for an automated system that can not only detect and track but also categorize fish as well as other aquatic species in underwater imagery, minimizing the manual interference. Absorption and scattering of light in deep sea environment leads to low resolution images making fish species recognition and classification a challenging task. Further, performance of traditional computer vision techniques tends to degrade in underwater conditions due to the presence of high background clutter and highly indistinct features of marine species. For such classification problems, Artificial Neural Networks (ANN) or deep neural network are being increasingly employed for improved performance. But the limited dataset of fish images makes it difficult to train such networks as they require huge datasets for training. Thus to reduce the requirement for a huge amount of training data, an algorithm using cross convolutional layer pooling on a pre-trained Convolutional Neural Networks (CNN) is proposed. The present paper focuses on the development of automatic system for classification, which can detect and classify fish from underwater images captured through videos. Thorough analysis on image dataset of 27,370 fish images gives a validation accuracy of 98.03%. The proposed method will be an efficient replacement to strenuous and time consuming method of manual recognition by marine experts and thus be advantageous for monitoring fish biodiversity in their natural habitats.