FISH RECOGNITION USING SUPPORT VECTOR MACHINE

Abstract

Fish image recognition has featured prominently as a core research area in recent years. Challenges currently militating against accurate fish recognition include distortion, noise, segmentation errors, overlap and occlusion of the fish images. Several techniques including Principal Component Analysis (PCA), K-Nearest Neigbour (K-NN), classification tree, and recently Neural Network had been widely used to resolve these challenges. Each of these approaches has inherent limitation which limits their recognition accuracy. In this research, a machine leaning technique called Support Vector Machine (SVM) is proposed for classification of fish species. A fish recognition model was built using SVM which uses two datasets (shape feature dataset and image texture dataset) which comprising of two different fish species (Ethmalosa fibriata and Scomberomorus tritor), with 75 Ethmalosa fibriata fish and 75 Scomberomorus tritor fish making a total of 150 fishes. The shape features dataset was built in collaboration with the department of Fisheries, FUTA and AAUA by making available Fisheries books and materials where the shape features of the fish were manually extracted. The shape feature dataset consists of six shape features (body length, anal fin length, caudal fin length, dorsal fin length, pelvic fin length and pectoral fin length). Also, the image texture dataset was acquired by taking the digital photo of 75 Ethmalosa fibriata fish and 75 Scomberomorus tritor fish making a total of 150 fish image. During the implementation stage, the dataset was divided into 2 which consist of 76 fish as training set and 74 fish as test set. The classification accuracy of separating Ethmalosa fibrata species from Scomberomorus tritor was found to be 74.32% based on shape features, while the classification accuracy was 78.59% based on image texture. A comparison of the obtained result when using SVM with results of some other classifiers such as ANN, K-NN, K-means Clustering and PCA shows that SVM has a higher degree of accuracy in recognition of fish species.