A Study on Feature Extraction of Handwriting Data Using Kernel Method-Based Autoencoder

Abstract

We use kernel method-based autoencoder in feature extraction application and evaluate its performance with a public handwriting database. Neural network-based autoencoder is an unsupervised algorithm and model that tries to learn an approximation function so as to extract features from data. Kernel method-based autoencoder has the same function compared with neural network-based autoencoder, but uses kernel methods to implement linear and non-linear data transformation. We use a handwriting dataset to evaluate kernel-based autoencoder, and examine the result by mean square error estimator, structural similarity index and peak signal-to-noise ratio for measuring image quality. We also investigate parameters of kernel functions to observe changes in the performance of the autoencoder. We found that effectiveness of kernel method-based autoencoder depends on the selection of kernel function and its parameter.