In this paper

we present a multiscale texture image segmentation algorithm based on adaptive window fixing and propagation.After raw segmentation based on wavelet domain hidden markov tree model

we use the fine consistency in coarse scale and accurate edges in fine scale of the raw segmentation result

label the homogeneity region and discontinuity region with different markers according to the characteristics of Markov between the node and its neighborhood and its father node

then we fix the homogeneity region and transfer their labels into their childrens;To discontinuity region

we determine a weight context by the neighbor information and transfer into the next scale to instruct the segm

entation with the information of this scale

finally we maintain the uniformity of regions and accuracy of edges.At the same time

we use an algorithm based on the Local polynomial approximation and Intersection of Confidence Intervals(LPA

-

ICI) to find out the physical location of centers of each texture due to the feature of aggregation of textures

then integrate them into the weight context so as to lead better segmentation.The experimental results shown that both on segmentation accuracy and boundary localization are greatly improved to synthetic texture images

and the proposed method is fast and does not need training.