关键词: 小波域, HMT模型, 图像分割, 上下文融合

Abstract: The wavelet-domain hidden Markov tree (HMT) model can capture multiscale texture features of objects in remote sensing imagery to obtain high quality segmentation. However, the existed wavelet-domain HMT segmentation methods directly fuse information of dyadic squares that are with different sizes at adjacent scales, resulting in the failure to get detail-preserving segmentation results. Therefore, a wavelet-domain HMT model based on iterative context fusion for image segmentation is proposed. In the proposed algorithm, context-based interscale information fusion is carried out in an iterative way. The current fused result at a scale is treated as the contextual information of the current scale to be refused in each loop, and the procedure stops when the difference between two successive iterations is smaller than the specified change threshold. Segmentation experiments are done using a synthetic image composed of Brodatz textures and a remote sensing image from Formosat-2 satellite, and quantitative and qualitative evaluation demonstrates that the proposed algorithm can preserve details better than traditional methods and achieve better segmentation performance.

Key words: wavelet-domain, HMT model, image segmentation, context-based fusion