Open Collections

Feedforward neural network design with application to image subsampling

University of British Columbia

Feedforward artificial neural networks (FANNs), which have been successfully applied to various image processing tasks, are particularly suitable for image subsampling due to their high processing speed. However, the performance of FANNs in image subsampling, which depends on both the FANN topology and the FANN training algorithm, has not been acceptable so far. High performance image subsampling is important in many systems, such as subband decomposition systems, and scalable image and video processing systems. This thesis addresses the design of FANNs with application to image subsampling. More specifically, we focus on both the topological design of FANNs and the training algorithm, so that efficient FANN structures, yielding good performance in image subsampling, are obtained. That is, we aim at obtaining compact FANNs that yield good subsampled versions of the original images, such that if reconstructed, they are as close as possible to the original images. Moreover, we aim at obtaining better performance-speed tradeoffs than those of the traditional lowpass filtering and subsampling methods. First, we propose a design method for FANNs, which leads to compact tridiagonally symmetrical feedforward neural networks (TS—FANNs). Next, in order to address the problem of artifacts that generally appear in the reconstructed images after FANN-based subsampling, we propose a training method for FANNs. When applied to first-order (FOS) and multi-stage first-order (MFOS) image subsampling, the FANNs trained using our method outperform the traditional lowpass filtering and subsampling (LPFS) method, without requiring pre- or post-processing stages. Motivated by our observation that the computational demands of the MFOS process increase approximately linearly with the image size, we then combine the proposed methods and evaluate the performance-complexity tradeoffs of the resulting TS-FANNs in FOS and MFOS. We show that our TS-FANNs-based subsampling has important advantages over subsampling methods based on fully connected FANNs (FC—FANNs) and LPFS, such as significantly reduced computational demands, and the same, or better, quality of the resulting images. The main contributions of this thesis consist of a method for FANN design with tridiagonal symmetry constraints, a training algorithm for FANNs applied to image subsampling, the design and evaluation of the performance-speed tradeoffs of FC—FANNs in image subsampling, and the design and evaluation of the performancespeed tradeoffs of TS—FANNs in image subsampling. The FANN performance in image subsampling is evaluated objectively (using the peak signal-to-noise ratios), subjectively (by visual examination of the subsampled and of the reconstructed images), and in the context of a video coding application. The speed and memory demands of the designed FANN structures are evaluated in terms of the subsampling time and the number of FANN parameters, respectively.

Text

2012-03-15

For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use.

http://hdl.handle.net/2429/41445

Electrical and Computer Engineering

University of British Columbia

Graduate