Applying MLP and CNN on Handwriting Images for Image Classification Task

Abstract

Nowadays, image recognition technology has been widely used in various fields, and some models like Multi-layer perceptron (MLP) and Convolutional Neural Network (CNN) both have advantage in processing image classification. The study compares the performance of these two common models and analyzes their data scale sensitivity when dealing with image classification on handwriting images. Based on the MNIST database, this study divides the training figures into 30000, 40000 and 60000 images and optimizes MLP by selecting proper activation function and increasing the number of dropout layers, neurons and hidden layers. Finally, the best performance of MLP is obtained when the model has three hidden layers, the number of neurons in each hidden layer is 1300, and the activation function for each layer is ReLU, ReLU and sigmoid respectively. The study also performs the model performance comparison with three different data volumes, and the best MLP model can achieve the accuracies of 97.49%, 97.98% and 98.24% separately when using 30000, 40000 and 60000 training images. Besides, the training curve and classification analysis of MLP and CNN are attained from the study. The results show the effects and the analysis of data quantity sensitivity of MLP and CNN for processing handwriting images, which provides a good reference for the future study of MLP and CNN when tackling with image classification and data sensitivity analysis.