Time: Monday, February 27, 2017, 1:00pm @ MTH3206

Speaker: Dr. Michelle Quirk (Los Alamos)

Title: Principal Component Filter Banks: Optimality Criteria for Multi-Dimensional Signal Encoding

Abstract: This talk is concerned with optimality criteria for multi-dimensional signal encoding, with applications to extensions to the JPEG2000 Standard (multi-component imagery compression.) Multi-dimensional signals, regarded as stochastic processes, are subject to pre-processing for an efficient encoding or transmission. The essence of processing prior to operations, such as compression or brute-force dimensionality reduction, is the de-correlation of the signal, or a packing of the energy in fewer dimensions of interest, following explicitly-stated criteria.

This talk begins with a brief on filter banks, the Karhunen-Loeve transform, and optimality criteria for orthonormal filter banks. Criteria for efficient de-correlation and encoding are next, followed by a discussion on the existence of finite impulse response - or realizable - optimal filter banks.

We conclude this talk with the design of two-channel filter banks for remote sensing hyper-spectral images, and results on feature extraction from compressed/reconstructed images, using the JPEG 2000 Standard engine.