Hierarchical Classification System with Reject Option for Live Fish Recognition

Abstract

This chapter presents a Balance-Guaranteed Optimized Tree with Reject option (BGOTR) for live fish recognition in a non-constrained environment. It recognizes the top 15 common species of fish and detects new species in an unrestricted natural environment recorded by underwater cameras. This system can assist ecological surveillance research, e.g., obtaining fish population statistics from the open sea. BGOTR is automatically constructed based on inter-class similarities. We apply a Gaussian Mixture Model (GMM) and Bayes rule as a reject option after hierarchical classification-we estimate the posterior probability of being a certain species and then filter out less confident decisions. The proposed BGOTR-based hierarchical classification method achieves significant improvements compared to state-of-theart techniques on a live fish image dataset of 24,150 manually labeled images from the south Taiwan sea. IntroductionLive fish recognition in the open sea has been investigated to help understand the marine ecosystem, which is vital for studying the marine environments and promoting commercial applications. This recognition task is fundamentally challenging because of its complex situation where the illumination changes frequently. Prior research is mainly restricted to constrained environments (fish in the tank or on a conveyor system) or dead fish, and these machine vision systems have only explored applications for a limited number of fish species. These methods perform worse when they deal with unconstrained fish in a real-world underwater environment, especially when the dataset is greatly imbalanced.

A Robust Image Enhancement Techniques for Underwater Fish Classification in Marine Environment

“…This study implements the novel contrast-adaptive color correction (NCACC) technique proposed by the authors in the previous work [7] to enhance image data, which are then transformed the image into useful features such as angular second moment, contrast, correlation, homogeneity, and entropy. These features are obtained by using the gray-level co-occurrence matrix (GLCM) method [28], and the image data are then classified using the backpropagation neural network classification method (BPNN).…”

“…The characteristics of the image are critical components of human visual perception, which are obtained based on color, shape, and texture. One method that can be used to do this is to use gray-level co-occurrence matrix (GLCM) [22], [28].…”

“…The orientation directions in computational GLCM are 0 0 , 45 0 , 90 0 , and 135 0 . The results of the GLCM matrix are the co-occurrence matrix and it can be used to obtain angular secondmoment features, contrast, correlation, homogeneity, and entropy [22,28].…”