PhD defence of Pavel Sinha – Properties and Applications of a Structurally Regularized CNN Architecture via Adaptive Subband Decomposition

Abstract

In this thesis, we present the concept of subband decomposition CNN architecture that can learn the subband decomposition coefficients from a training dataset through the automated end-to-end training framework of a neural network. We present a few variations of the architecture that provide trade-off between computational cost and performance. Further, we analyze the properties of the subband decomposition architectures and compare their performance with present state-of-the-art architectures. We observe the subband decomposition CNN architecture to perform at part of the state-of-the-art CNN architectures at a fraction of the computation cost. We then apply the concept of subband decomposed CNN architecture to the problem of image segmentation of lumen and media in IVUS images. We compare results with state-of-the-art CNN architectures and easily identify that not only does the subband decomposition architecture outperform the competition, but does that at a fraction of the computation cost.

Further, we apply the concept of subband decomposed CNN architectures to the problem of channel equalization and synchronization in OFDM radio receiver subsystems. To understand the practicability of the subband decomposed CNN architecture-based receiver synchronization subsystem, we implement it on an FPGA and analyze the hardware resource utilization for real-life applicability.