Image Recognition Using Invariant Moments

Abstract

This thesis investigates the use of invariant moment descriptors for image recognition, namely Hu, Zernike, Pseudo-Zernike, Legendre, and Tchebichef moments. The study aims to evaluate their ability to represent image shapes and improve classification performance across several datasets, including MNIST, MPEG-7, FLAVIA, CIFAR-10, and others. After preprocessing, moment-based features were extracted and classified using the k-Nearest Neighbors (k-NN) classifier. The experimental results show that invariant moments are effective for image recognition, especially in shape-based tasks. Legendre moments achieved the best overall performance across most datasets, followed by Zernike and Pseudo-Zernike moments, while Hu moments generally produced lower accuracy. These findings confirm the usefulness of invariant moments in computer vision applications.