CAE-CNN: Image Classification Using Convolutional AutoEncoder Pre-Training

Abstract

The work presented in this paper is in the general framework of classification using deep learning and, more precisely, that of convolutional neural networks (CNN). In particular, the convolutional autoencoder proposes an alternative for the processing of high-dimensional data, to facilitate their classification. In this paper, we propose the incorporation of convolutional autoencoders as a general unsupervised learning data dimension reduction method for creating robust and compressed feature representations in order to improve CNN performance on image classification tasks. For prediction reasons, we applied the two methods to the MNIST image databases. The use of CNN with the convolutional autoencoder gives better results compared to the individual use of each of them in terms of accuracy, to obtain a good classification of the data high-dimensional entrance.