JPEG-like Image Compression using Neural-network-based Block Classification and Adaptive Reordering of Transform Coefficients

Grosse, Hanns-Juergen (1997) JPEG-like Image Compression using Neural-network-based Block Classification and Adaptive Reordering of Transform Coefficients. Doctoral thesis, University of Central Lancashire.

[thumbnail of Thesis document]
Preview
PDF (Thesis document) - Accepted Version
Available under License Creative Commons Attribution Non-commercial Share Alike.

5MB
[thumbnail of Permission form] PDF (Permission form) - Supplemental Material
Restricted to Repository staff only

466kB

Abstract

The research described in this thesis addresses aspects of coding of discrete-cosinetransform (DCT) coefficients, that are present in a variety of transform-based digital-image-compression schemes such as JPEG. Coefficient reordering; that directly affects the symbol statistics for entropy coding, and therefore the effectiveness of entropy coding; is investigated. Adaptive zigzag reordering, a novel versatile technique that achieves efficient reordering by processing variable-size rectangular sub-blocks of coefficients, is developed. Classification of blocks of DCT coefficients using an artificial neural network (ANN) prior to adaptive zigzag reordering is also considered.

Some established digital-image-compression techniques are reviewed, and the JPEG standard for the DCT-based method is studied in more detail. An introduction to artificial neural networks is provided.

Lossless conversion of blocks of coefficients using adaptive zigzag reordering is investigated, and experimental results are presented. A versatile algorithm, that generates zigzag scan paths for sub-blocks of any dimensions using a binary decision tree, is developed. An implementation of the algorithm based on programmable logic devices (PLDs) is described demonstrating the feasibility of hardware implementations. Coding of the sub-block dimensions, that need to be retained in order to reconstruct a sub-block during decoding, based on the scan-path length is developed.

Lossy conversion of blocks of coefficients is also considered, and experimental results are presented. A two-layer feedforward artificial neural network trained using an error-backpropagation algorithm, that determines the sub-block dimensions, is described. Isolated nonzero coefficients of small significance are discarded in some blocks, and therefore smaller sub-blocks are generated.


Repository Staff Only: item control page