Patch Autocorrelation Features for Optical Character Recognition

Abstract

The autocorrelation is often used in signal processing as a tool for finding repeating patterns in a signal. In image processing, there are various image analysis techniques that use the autocorrelation of an image for a broad range of applications from texture analysis to grain density estimation. In this paper, a novel approach of capturing the autocorrelation of an image is proposed. More precisely, the autocorrelation is recorded in a set of features obtained by comparing pairs of patches from an image. Each feature stores the euclidean distance between a particular pair of patches. Although patches contain contextual information and have advantages in terms of generalization, most of the patch-based techinques used in image processing are heavy to compute with current machines. Therefore, patches are selected using a dense grid over the image to reduce the number of features. This approach is termed Patch Autocorrelation Features (PAF). The proposed approach is evaluated in a series of handwritten digit recognition experiments using the popular MNIST data set. The Patch Autocorrelation Features are compared with the euclidean distance using two classification systems, namely the k-Nearest Neighbors and Support Vector Machines. The empirical results show that the feature map proposed in this work is always better than a feature representation based on raw pixel values, in terms of accuracy. Furthermore, the results obtained with PAF are comparable to other state of the art methods.