Abstract: Thin details in the filtered images are suppressed by the probabilistic patch-based (PPB) filter,which is attributed to the absence of effective selection of pixel patches and the unsuitable method of weight computing．For these problems,the data structure of cluster tree was introduced firstly．The same distance measure as applied in the PPB filter was chosen to build the cluster tree,which allows for efficient and precise selection of similar patches．Since the origi-nal PPB filter could not handle rotated or mirrored repetitive regions properly,the weight between two patches was redefined after the rotation of the patches．Finally,the PPB (non-it) filter was used for the denoising．Experimental results show that the improved filter has better performance in texture and details preservation than the original PPB (non-it) filter,especially in retaining thin details．

Key words: SAR image,Despeckling,Clustering,PPB (probabilistic patch-based) filter,Cluster tree