Classification of fish images based on shape characteristic

Abstract

This research work has been conducted to analyze and classify the types of fish image based on shape characteristic. The features of characteristic of fish image are extracted by using three Moment Invariants (MI) techniques and Fourier descriptors (FD). The types of Moment invariants are Geometric moment invariant (GMI), United moment invariant (UMI), Zernike moment invariant (ZMI). In the FD’s technique, there are two edge detection have been used to create the boundary of the image, namely Robert cross detection and Sobel cross detection. These feature extraction techniques have been used to analyze the image due to its invariant features of an image based on translation, scaling factor and rotation. There are two ways to examine the performance of feature extraction techniques, namely intra-class analysis and classification analysis. For the intra-class analysis, a set of equations has been implemented to find the best technique among the three different types of moments and Fourier descriptors based on the low value of Total Percentage Min Absolute Error (TPMAE). Meanwhile, for the classification analysis, the Artificial Neural Network (ANN) is explored and adapted to classify the fish images. The feature vectors produce by feature extraction techniques that represent the image are used as the input of classification. The results of the intraclass analysis indicate that the UMI was the best technique among the moment techniques while Fourier descriptor by using the Sobel edge detection shows the lower TPMAE as compared to Robert edge detection. For the classification part, two types of ANN’s which are Multilayer Perceptron (MLP) and Simplified Fuzzy ARTMAP (SFAM) neural networks have been used to classify the image based on fish category. The Leverberg-Marquardt (LM) algorithm is used to train the MLP network in order to check the applicability. Based on the classification that has been computed, the results show that all networks perform good classification performance with overall accuracy is around 90%. However, the MLP trained by Leverberg-Marquardt shows the highest classification performance in classifying the fish images as compared to the SFAM network.