An efficient approach for dynamic PCA filter selection in PCANet for image classification

Abstract

Character recognition is capturing incremental interests due to the revoked interests in text recognition. Optical character recognition converts handwritten images into machine encoded text. For character recognition, feature extraction and its representation are important issues, which has been pursued in recent years. Principal Component Analysis Network (PCANet) is a very simple deep learning network for image classification designed by cascading PCA filters same as convolution filters in convolution neural network. PCANet uses very basic data processing components: cascaded principal component analysis (PCA), binary hashing and block wise histograms. In this paper, we apply gradient descent to learn the number of filters used in the various stages of PCANet. The proposed method achieves promising performance on the Modified National Institute of Standards and Technology database (MNIST) dataset and on the Devanagari handwritten character dataset (achieving an accuracy of 98.946% and 84.09% respectively), demonstrating the effectiveness of learning the number of convolution filters in PCANet using gradient descend in character recognition.