Efficient neural network training using curvelet features

Sparse representations for signals and images have been used extensively in various image processing tasks. In this work, we use the curvelet transform as a sparsity inducing tool in neural networks. Nowadays, there is much interest in research and development of efficient algorithms that reduce the computational demands of training neural networks. We demonstrate that the compact directional representation generated by the curvelet transform has the potential of significantly reducing computational time and costs. In our proposed classification framework, curvelet coefficients (amplitude and angle) are mapped into the complex domain. Next, a single-layer Complex-Valued Neural Network (CVNN) is used to represent the complex input parameters. After training the network using a suitable training dataset, computed complex weight vectors are used to classify testing datasets. The developed algorithm was tested on the MNIST dataset. Our results indicate a performance that is comparable with state-of-the-art while offering dramatic computational savings.