The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical and Computer Engineering University of Southern California

Technical Report USC-SIPI-114

“Textured Image Segmentation Using Feature Smoothing and Probabilistic Relaxation Techniques”

by John Y.-H. Hsiao

December 1987

The segmentation of textured imagery into homogeneous regions is an important and difficult task in scene analysis. A satisfactory segmentation result should possess not only good region interiors but also accurate boundaries. In response to these requirements, the main objective of this research is twofold. First, to improve textured image segmentation results; especially along the borders of regions. Second, to take into account the spatial relationship among pixels to improve the segmentation of region interiors.

An improved method to extract textured energy features in the feature extraction stage is proposed. The proposed method is based upon an adaptive noise smoothing concept which takes the nonstationary nature of the problem into account. Texture energy features are first estimated using a window of small size to reduce the possibility of mixing statistics along region borders. The estimated texture energy feature values are then smoothed by a quadrant method which reduces the variability of the estimates while retaining the region border accuracy.

For a supervised segmentation system, the estimated feature values of each pixel are used by the Bayes classifier to make an initial probabilistic labeling. The spatial constraints are then enforced through the use of a probabilistic relaxation algorithm. Two probabilistic relaxation algorithms have been investigated and their results are compared by computer simulation.

For an unsupervised segmentation system, the estimated feature values of a set of subsampled pixels are used in a K-means clustering algorithm to estimate the class statistics. The estimated class statistics are then used by the Bayes classifier to make an initial probabilistic labeling. One of the selected relaxation algorithms is then applied to enforce the spatial constraints.

Limiting the probability labels by probability threshold is proposed to make the relaxation iteration more efficient. The trade-off between efficiency and degradation of performance is studied. Finally, an overview of the proposed textured image segmentation system is provided and comparisons of overall performance are made.

To download the report in PDF format click here: USC-SIPI-114.pdf (13.1Mb)