Initial parameters of CNNs generated by Convolutional Sparse Representation with L1 error term

Open Access

Abstract: Convolutional Sparse Representation (CSR) approximates images with the convolutional sum of dictionary filters and corresponding sparse coefficients. To improve classification accuracy of Convolutional Neural Networks (CNNs), this paper proposes to use the dictionary filters generated by CSR as initial parameters of CNNs’ filters since the CSR filters express features of test images. Our method also estimates the error term of CSR with the L1 norm instead of the L2 norm to increase robustness against outliers in datasets for training. The results of experiments classifying CIFAR-10 show that the CNN using the initial parameters generated by the proposed method with the L1 error term shows the highest classification accuracy for small numbers of training images compared with the two methods: the proposed method with the L2 error term and the Xavier’s method.

Keywords: Convolutional Sparse Representation, Convolutional Neural Network