Classification Performance Analysis of MNIST Dataset Utilizing A Multi-Resolution Technique

Abstract

Here, we propose a method for recognition of handwritten English digit utilizing discrete cosine space-frequency transform known as the Discrete Cosine S-Transform (DCST). Experiments have been conducted on the publicly available standard MNIST handwritten digit database. The DCST features along with an Artificial Neural Network (ANN) classifier is utilized for solving the classification issues of written by hand digit. The Discrete Cosine S-Transform coefficients are extracted from the standard images of MNIST handwritten isolated digit database. The database consists of a total of 70000 including 60000 training samples and 10000 test samples. To overcome the computational overhead, we have normalized all the images of the MNIST dataset from 28 28 to 20 20 image size by eliminating the unsought boundary pixels up to width four. Further, the classification of digits has been made by using a back propagation neural network (BPNN). This work has achieved precisely 98.8% of success rate for MNIST database.