|
Paper Abstract:
|
We propose a new supervised texture segmentation and classification technique based on combining features extracted
from the discrete wavelet frames of an image (specifically, the detail images at multiple resolutions) with a nonlinear
band generation algorithm and an orthogonal subspace projection operator (OSP). As a supervised technique, the
algorithm needs apriori information about the number and location of textures present in the composite texture training
images. The OSP operator role is twofold: to extract a set of texture signature vectors each uniquely characterizing only
one texture; after that, the texture segmentation process commences and the signature vectors are used to identify/mark
textures in new images, essentially a pixel labeling process with all pixels of one texture having the same label. The
simulation results show satisfactory classification and segmentation on a set of composite texture images while having
good real time performance and moderate storage and computational requirements. |
