Deep Neural Network Initialization Methods for Micro-Doppler Classification With Low Training Sample Support

Motion Classification Using Kinematically Sifted ACGAN-Synthesized Radar Micro-Doppler Signatures

“…To analyze the improvement achieved by the generative Read real spectrograms 3: Apply GPCA subspace learning method and find the optimized subspaces, (100x100 spectrogram dimensions reduced to 2x2) 4: Find the boundaries of the reduced feature space using the 4 dimensional convex hull method 5: Save the convex hull parameters and optimized subspaces. 6: end for 7: for EACH CLASS do 8: Read the genered spectograms 9: Load the optimized subspaces and convex hull parameters 10: Apply optimized subspaces and get the reduced feature space 11: Check if the current generated spectrogram is within the convex hull boundaries with a tolerance 12: If in keep else eliminate 13: end for 14: return Sifted images In prior studies, VGGnet is a network that has provided accuracies that have surpassed that of other pretrained networks, such as GoogleNet, as well as convolutional autoencoders (CAEs) and supervised learning with handcrafted features [56], when the amount of training data is limited [19]. Therefore, it is not suprising that VGGnet surpasses the performance of AlexNet, and even that attained by using the unsifted, initial training database generated by ACGAN without consideration of any kinematics.…”

“…Deep neural networks (DNNs) have shown great potential to achieve high accuracy, even as the number of classes increases, and may well lead the way as a preferred method for motion classification in the near future [14]- [19]. Yet, DNN architectures in RF applications are often limited by the fact that only small datasets are available for training, as data acquisition can be time consuming, costly, and limited in terms of the scope of scenarios and targets sampled.…”

DopNet: A Deep Convolutional Neural Network to Recognize Armed and Unarmed Human Targets

“…Since Mdist is designed to ensure the discrimination capability of the matrix, the regularized term denoted as L M E , is added to balance the E-distance and M-dist, which enforces the term w T f c2 w f c2 to be close to identity matrix, as shown in Eqn. (8). By adjusting the balance the E and M distance, we are actually controlling the discrimination and generalization of the DCNN.…”

“…In addition, µ-DS have been used to distinguish wind turbine blades and the blades of aircraft rotors [7]. It has also been demonstrated how µ-DS of different human target movements can help increase the situational awareness of the ambient assistant living in the healthcare context [8][9][10][11][12][13][14].…”

“…Recently, DCNNs have also been used for classification task of aquatic movement using the fine-tuning method, which utilized the trained DCNN network weights by ImageNet dataset so that only small part of the network weights are required to be trained [26]. For healthcare applications, this idea is further evaluated and compared with the one using auto-encoder based pre-training weights in the work of [8], which handles fine-tuning DCNNs with deep layers (e.g. VGG-Net and Inception-Net) using a small number of radar samples.…”