Deep neural network initialization and training methodologies for radar micro doppler signature classification

Recently, deep neural networks (DNNs) have been the subject of intense research for the classification of radio frequency (RF) signals, such as synthetic aperture radar (SAR) imagery or micro-Doppler based automatic target recognition. However, a fundamental challenge is the typically small amount of data available due to the high costs and resources required for measurements. Deep neural networks (DNNs), however, require large-scale labeled datasets to prevent overfitting while having good generalization. In this thesis, the efficacy of two neural network initialization techniques - unsupervised pre-training and transfer learning - for dealing with training DNNs on small data sets are compared. Unsupervised pre-training is implemented through the design of a convolutional autoencoder (CAE), while transfer learning from two popular CNN architectures (VGGNet, GoogleNet and ResNet). Also, a novel method for generating diversified radar micro-Doppler signatures using Kinect-based motion capture simulations is proposed as a training database for DNNs exploiting transfer learning. Results have also been compared with the conventional approach where handcrafted features are employed to train an SVM classifier. In particular, it is shown that the proposed residual transfer learning network, DivNet, offers the highest classification accuracy among other DNN architectures and, is robust to image noise.