Texture segmentation and characterization using Adaptive Gram-Schmidt modelling

Textured images pose considerable difficulties when processed with traditional image processing techniques. This paper describes the application of a texture characterization algorithm based on modelling inter-pixel correlations using linear prediction filters in different directions. The basis of the algorithm is the prediction of a pixel grey-level value from a group of surrounding pixels using locally optimum modelling parameters based on the Least Mean Squares (LMS) criterion. Previous work on using the LMS predictor proved useful in texture classification and segmentation, despite its slow convergence. This paper will describe the use of an afternative realization, which employs an adaptive Gram Schmidt pre-processor to decorrelate pixel values, thus achieving faster convergence. The algorithm is suitable for parallel implementation. In this paper, the modified algorithm, which consists of four predictors, is described along with the post-processing stage to segment and classify textures. Details of the design of the Adaptive Gram Schmidt (AGS) predictor are also given. Comparative results for the direct and AGS algorithm on both artificial and real images will be presented and discussed. The description also includes information on how the presented algorithms can be realised on a distributed system to speed up computation.