In this paper, we present a new approach for designing filter banks for ima
ge compression. This approach has two major components: optimization and ge
neralization. In the optimization phase, we formulate the design problem as
a nonlinear optimization problem whose objective consists of both the perf
ormance metrics of the image coder, such as the peak signal-to-noise ratio
(PSNR), and those of individual filters. Filter banks are optimized in the
optimization phase based on a set of training images. In the generalization
phase, the filter bank that can be generalized to other images is selected
from the candidates obtained in the optimization phase to be the final res
ult. The filter bank selected should perform well not only on the training
examples used in the design process but also on test cases not seen. In con
trast to existing methods that design filter banks independently from the o
ther operations in an image compression algorithm, our approach allows us t
o find filter banks that work best in a specific image compression algorith
m for a certain class of images. In system prototype development, we adopt
the agent-based approach to achieve better modularity, portability, and sca
lability. Agents in the multi-agent system are specialized in performing pr
oblem formulation, image compression, optimization, and generalization. In
the experiments, we show that on a set of benchmark images our approach has
found filter banks that perform better than the existing filter banks in d
ifferent image compression algorithms and at different compression ratios.
(C) 2001 Published by Elsevier Science Inc.