Many of the numerous texture measurements are based on space-frequency sign
al decomposition; these include Gabor filters and wavelet-based methods. Th
e discrete cosine transformation (DCT) extracts spatial-frequency (SF) comp
onents from a local image region. It is the basis for the JPEG image compre
ssion standard and has many fast algorithmic implementations. By using a sl
iding DCT we derive a SF representation for a region of interest (ROI) surr
ounding each image pixel. We show that the DCT coefficients may represent a
SF as a combination of several DCT coefficients depending on the offset of
the SF waveform maximum from the ROI's beginning, Thus, the DCT coefficien
ts for a texture with a certain SF will change as the transformation is mov
ed over the texture. In order to circumvent this problem, we derive horizon
tal and vertical SF shift-insensitive measurements from DCT components, Exa
mples are given which show how these DCT shift-insensitive (DCTSIS) descrip
tors can be used to classify textured image regions. Since a large number o
f image display, storage and analysis systems are based on DCT hardware and
software, DCTSIS descriptors may be easily integrated into existing techno
logy and highly useful. (C) 2000 Pattern Recognition Society. Published by
Elsevier Science Ltd. All rights reserved.