Improving Radar Image Classification using Dictionary Learning for Filter Initialization in Convolutional Neural Networks

Abstract:
The lack of data is a common problem in radar based neural network applications that require a labeled dataset. An unsupervised data analysis method might be used to initialize the network parameter. In this paper, the dictionary learning technique, which is well known in the compressed sensing community, is used to initialize the filter in the first layer of a convolutional neural network. Three different dictionary learning algorithms, namely K-SVD, ODL and DOMINODL, are used to analyze the MSTAR dataset and to create filters for the first network layer. The results show an improved training performance, i.e. a faster convergence of the network, and in some cases also a better classification of the test data with small sized networks. In this way, the training time, which is a common bottleneck in neural network application, and the number of parameters, i.e. weights, of a neural network can be reduced.