Nonlocal Means SAR Image Despeckling using Principle
Neighborhood Dictionaries
Hua Zhong *, Chen Yang and Lc Jiao
Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education,
Xidian University, Xi’an 710071, China.
ABSTRACT
The Principle Neighborhood Dictionary (PND) filter projects the image patches onto a lower dimensional subspace using
Principle Component analysis (PCA), based on which the similarity measure of image patch can be computed with a
higher accuracy for the nonlocal means (NLM) algorithm. In this paper, a new PND filter for synthetic aperture radar
(SAR) image despeckling is presented, in which a new distance that adapts to the multiplicative speckle noise is derived.
Compared with the commonly used Euclidean distance in NLM, the new distance measure improves the accuracy of the
similarity measure of speckled patches in SAR images. The proposed method is validated on simulated and real SAR
images through comparisons with other classical despeckling methods.
Keywords: PND, nonlocal means, despeckling, SAR
1. INTRODUCTION
Synthetic Aperture Radar (SAR) image is widely applied in military and civilian, but the effect of inherent speckle leads
to that it can not effectively reflects the scattering properties of targets, so the speckle seriously influence the quality of
the image. However, SAR image despeckling remains an unsolved problem, which aims to remove speckle noise while
retain image features such as texture, edges, point-type targets, etc. Spatial domain based methods such as Lee filter [1],
Kuan filter [2], Frost filter [3] and Gamma MAP filter [4], approach the local mean at homogeneous regions, while tend
to retain the original observation at pixels of high activity. Their disadvantage is either oversmoothing the image texture
or ineffective denoising around edges. WBE filter, which is proposed for SAR image in [9], can effectively keep the
mean of image. When developing despeckling method, researchers introduced the ideas of many image denoising
methods for additive Gaussian noise, for example, wavelet based methods [5], model based methods [6], etc.
The nonlocal means (NLM) filter was first proposed by Baudes et al. [7] for additive Gaussian noise reduction. An
estimator of the central pixel can be acquired by the weighted average of all pixels in a given image, while the each
weight is based on the Euclidean distance computed between the patch that centre the central pixel and the one that
centre the neighboring pixel of image. Due to its great power in noise removal while detail preservation, the NLM based
methods has become one of focus in image denoising field.
As a recent approach, Principle Neighborhood Dictionary (PND) filter introduced by Tolga Tasdizen [8] projects the
image patches onto a lower dimensional subspace using PCA, based on which the similarity measure of image patch can
be computed with a higher accuracy for the NLM algorithm. Besides, the subspace dimensional selection is attained by a
modification to the parallel analysis, which yields good result. The PND NLM filter performs better than the NLM filter
for additive Gaussian noise removal and computational cost. However, few methods are designed for SAR image
despeckling, since the commonly used Euclidean distance in NLM and the distance in PND NLM both are not fit for the
multiplicative speckle noise.
In this paper, we introduce a new PND NLM filter for SAR image despeckling. Our PND filter uses a new distance to
measure the similarity of speckled patch for SAR images, which can effectively reduce the impact of speckle and
different intensity. The new distance measure improves the accuracy of the similarity measure of speckled patches in
SAR images. Experimental results on synthetic and real SAR images demonstrate its effectiveness.
*hzhong@mail.xidian.edu.cn; phone 86 29 88204298; fax 86 29 88204298;