SAR Image Despeckling Based On Nonlocal Means Filtering

The special imaging mechanism of SAR leads speckle, which cause difficulties in the research of target recognition and feature extraction. Therefore, despeckling is an indispensable process. This dissertation studies the nonlocal means (NL-means) algorithm and applies it for SAR despeckling.A new NL-means despeckling method for SAR image is proposed, which is adapted to the multiplicative model of speckle noise. By this method, image pixels are first classified into different classes such as point, line, edge, surface, etc., using ratio edge detector. Then, different smooth parameters of NL-means filter are used according to the class information. In addition, a searching method for rotation-invariant similar patches is designed through the use of directional information, which improves the accuracy of similar patches searching.According to the disadvantages of ratio edge detector, a new detection template which has a fixed window size, uniform threshold, and more direction information is proposed. It can get better performance with smaller computational cost. Experiments prove that this algorithm can guarantee the coherence of edge and line, at the same time the amout of error line and edge targe is less compared with ratio edge detector.A new similarity measure formula is deduces for SAR image by researching the similarity measure formula of NL-means filtering. Real gray distance of pixels neighbor can be calculated by the observant gray value and the noise standard deviation. This algorithm overcome the problem between details maintain and smoothing degree Moreover, the algorithm is simple and easy to realize.The research is supported by NSFC(No.6050510,60702062).