SAR Image Despeckling Based on Combination of Laplace Mixture Distribution with Local Parameters and Multiscale Edge Detection in Lapped Transform Domain

摘要

The speckle noise badly affects the tasks of automatic information extraction and scene analysis in Synthetic Aperture Radar (SAR) images. Therefore, the despeckling of SAR images while preserving edge and textures is highly important. In this paper, a lapped transform (LT) based SAR image despeckling algorithm is proposed. Since lapped orthogonal transform (LOT) is orthogonal and has good energy compaction, the statistical modeling of noise and signal can be done precisely in the same domain. The use of LOT in denoising applications is motivated by its low computational complexity and its feature of robustness to over-smoothing. Since the LOT is block transform, the dyadic remapping is carried out first and then the subband LOT coefficients are modeled similar to wavelet coefficients. The LOT coefficients of the logarithmically transformed reflectance and the speckle noise are modeled using 2-state laplace mixture pdf that uses local parameters and zero mean Gaussian pdf respectively. A MAP estimator within Bayesian framework based on proposed prior is developed. The mixture distribution parameters are estimated using Expectation-Maximization (EM) algorithm. Subband LOT coefficients at each scale are classified into edge and non-edge coefficients using LOT modulus maxima computation. The non-edge coefficients are filtered using the proposed Bayesian MAP estimator and edge coefficients are kept unmodified. The method is implemented in ‘cycle spinning’ mode to solve the problem of lack of shift-invariance property of the LOT. Experimental results carried out on real SAR images show that the proposed scheme very effectively preserve the edges of a SAR image with notable speckle suppression and outperform two undecimated wavelet transform based methods.