Machine Learning Algorithm for Handwritten Numerical Character Identification Using FPGA Based Custom IP Core

Many recent advances in computer vision are based on the fundamentals of image classification techniques. These schemes have already been proved advantageous to the commer-cial, industrial, healthcare, and defense industries. Every pixel in image undergoes numerous mathematical operations during the classification processes. Sequential execution of these operations leads to a slower throughput. Field Programmable Gate Arrays (FPGAs) have been used to develop hardware accelerators for parallel image processing that shorten pixel processing time. The classification of handwritten letters is an important step in optical character recognition (OCR). The open source Modified National Institute of Standards and Technology's (MNIST) handwritten numerical digits dataset is frequently used to train handwritten character classification models. The goal of this work is to create a Multilayer Perceptron (MLP) based custom intellectual property (IP) core on the PYNQ-Z2 platform that can classify handwritten numerical characters with 97.48 % blind test accuracy.