Abstract: An HMTseg algorithm with the adaptive fusion mechanism is proposed for the segmentation of remotely sensed images. In this algorithm, the statistical correlation properties between wavelet coefficients across different scales are well employed. The initial segmentation results on coarse and fine scales are fused adaptively by the assignment of different contextual weights. This algorithm not only lays emphasis on the advantage of region consistency from the coarse-scale segmentation, thus preserving the main outlines of individual homogeneous regions, but takes into account the advantage of boundary location accuracy from the fine-scale segmentation, thereby favoring the discrimination of small targets within an image. Experimental results on remotely sensed images, including aerial photos and SAR images, demonstrate that the algorithm can improve the segmentation performance of remotely sensed images.

Key words: discrete wavelets transform, hidden Markov tree model, image segmentation, adaptive weighting, multiscale fusion