Network-conscious image compression
Sami Iren, Paul D. Amer · 1999
This dissertation proposes network-conscious image compression , a new perspective in image compression that aims not simply to optimize compression, but to optimize overall performance when compressed images are transmitted over a lossy, packet-switched network such as the Internet. Using an Application Level Framing philosophy, an image is compressed into path MTU-size Application Data Units (ADUs) at the application layer. Each ADU carries its semantics, that is, it contains enough information to be processed independently of all other ADUs. Therefore each ADU can be decoded and displayed by the receiving application out-of-order, enabling faster progressive image display. An integral component of this dissertation work has been the design and implementation of Network-Conscious Image Compression and Transmission System (NETCICATS), a software system for empirically investigating the combination of transport protocol features and network-conscious compression algorithms. Developed in the Protocol Engineering Laboratory of University of Delaware, NETCICATS coordinates software components from the network layer (e.g., lossy router), transport layer (e.g., ordered/unordered/partially ordered reliable transport service), and application layer (e.g., compression algorithms, browsers). The network-consciousness hypothesis was first tested by using an earlier version of NETCICATS that allows flexible control of image quality and size, compression algorithms and their parameter settings, and transmission parameters (e.g., QoS parameters, transport service, network loss rate). Initial experiments using wavelet transformed images motivated further investigation of the hypothesis. Two standard image compression algorithms, GIF and SPIHT, were modified to make them network-conscious: GIF-NC and SPIHT-NC. GIF-NC modifies GIF's file structure using an Application Level Framing Approach. SPIHT-NC uses the same wavelet zerotree encoding technique SPIHT uses, but produces an encoded bit stream that can be packetized such that each packet can be processed independently. Experimental results conclude that for certain QoS, network-conscious compressed images can perform better than standard images over lossy packet switched networks. In particular, network-consciousness can provide significant gain at higher loss rates, lower bandwidths, larger round trip delays, and larger window sizes.