HOG and Color Texture Salience: An Expedient Descriptor for Bangladeshi Fish Recognition

Abstract

The number of fish species is decreasing alarmingly. To protect them from extinction, they require a computer vision process to classify fish species and find the endangered ones. Much research is done on this topic and much more is going on. This study aimed to find a feasible process for detecting and classifying Bangladeshi fish species using image processing methods. The detection step works with the Histogram of Oriented Gradient (HOG) descriptor and various classifiers to differentiate between fish class and non-fish class. Four different classifiers: Support Vector Machines (SVM), Decision Tree, Random Forest, and Naïve Bayes classifier were attempted. A classification process using Color Co-occurrence Matrix (CCM) texture descriptor and SVM classifier is used to classify fish species. Some statistical properties such as contrast, dissimilarity, correlation were employed to constitute the feature vector from CCM. From experimentation, it was found that HOG with SVM offers a promising accuracy in the fish detection phase. Also in the classification stage, extracted statistical CCM features with an SVM classifier show a practically useful accuracy. This detection and classification process is a way to save those endangered fish species.