Fish Classification Using Saliency Detection Depending on Shape and Texture

Abstract

Classification of fishes becomes important after the advancement of machine learning. As fishes play a vital role in the economy of Bangladesh, a proper monitoring system will maximize the cultivation. It will also contribute to the overall economy. Therefore, here introduce a system that can detect the fishes and compare various methods with explanations to understand the selected methods. This paper have considered 5 categories of local fishes of Bangladesh in the dataset. The technique consists of preprocessing with segmentation, feature descriptor, and ensembles to produce the final result. U2 -net is used in the preprocessing layer to obtain two types of features namely shaped images and colored images with removed backgrounds. To get the features, we have used a histogram of oriented gradient (HOG) and an ensemble layer is used for classification purposes. Experimental results illustrate the accuracy of 99.77% for the first ensemble and 100% for the second ensemble layer on our dataset of 2678 fishes of 5 distinguishing classes. Various layers were used to compare the predicted results using different performance metrics.