Wavelet convolutional neural networks for handwritten digits recognition

Abstract

Image recognition is a very important subject in machine learning. With the increased use of intelligent applications, the image recognition has become more important in various domains. Thus, neural networks and deep learning algorithms has proved a notable success in the domain of image recognition and classification. Convolutional Neural Networks (CNN) are a class of deep learning methods. They are constrained from convolutional, pooling and fully connected layers. While they achieved great results in object recognition and classification, the pooling layer does not take into consideration the structure of the features. This paper emphasizes the pooling layer of CNN by adding a wavelet decomposition to obtain a new architecture called Wavelet Convolutional Neural Networks (WaveCNN). This architecture is validated on the handwritten digits recognition application using the MNIST benchmark. Compared to a conventional CNN with the same architecture, we found better results. Hence, WaveCNN is able to represent more adequately features for classification.