Algorithms for progressive image compression seek to represent an image in such a way that a small fraction of the data file can allow a displayed image to be recognizable, and a larger portion of the data allows improved image quality. The main applications of progressive image compression are for browsing of remote databases, Internet image hopping, and fast remote decision-making tasks where many images must be scanned. The goal is to compress images to rates so low as to yield "just-recognizable" quality. The user can then exploit just- recognizable images while progressively scanning an image database by "quitting" each image as soon as it is determined not to be the one sought. This research involves theoretical understanding of ultra-low rate progressive image coding both with and without channel noise, as well as the development and evaluation of practical algorithms for applications such as low-bandwidth video compression and packet network image transmission. One portion of this research focuses on algorithm development, including optimization of wavelet zerotree coding for ultra-low rates, and coding using regions or line caricatures. The investigators are studying the effect of channel noise on progressive coding systems, and are developing both grayscale and color bit progressive image coding algorithms that do not exhibit catastrophic breakdown in the presence of channel noise. One promising approach uses rate compatible punctured convolution codes applied to images compressed with a wavelet zerotree method. Experimental studies of response times for human observers of progressive image coders will allow the researchers to quantitatively compare one progressive image coder against another, and yield insight into human perceptual responses to just-recognizable imagery.
