Development of Efficient Speckle Suppression Algorithms for SAR Images

Abstract

Synthetic aperture radar (SAR) is an active imaging sensor that provides highresolution, cloud-free images in varying illumination and weather conditions. SAR images have been used in a wide range of applications such as remote sensing, geology, environmental monitoring, and military surveillance. However, these images are inherently noisy due to the presence of granular structures known as speckle. Its presence visually degrades the underlying image information and has an impact on subsequent image analysis. The amplitude based noisy SAR image follows Rayleigh statistics. This dissertation focuses on the development of efficient algorithms for speckle suppression in SAR images using variational methods. Among various speckle reduction filters, the variational methods restore the clean image by minimizing a suitable cost function. This function consists of a forward/data model derived from the speckle statistics and a prior/regularization model containing the information regarding the image. The minimization is carried out using a suitable optimization technique. Developing an effective prior is a non-trivial task, and the utilization of complex priors makes the optimization problem difficult to solve. The plug-and-play (PnP) is a flexible framework that substitutes priors with state-of-the-art denoisers and thereby simplifying the software integration. Further, it can be easily solved by adopting any appropriate splitting techniques, such as alternating direction method of multipliers (ADMM). The first contribution includes the continuation scheme of PnP ADMM for the reduction of Rayleigh based speckle in SAR images. The forward model is coupled with the standard denoisers, and the optimization problem is solved using ADMM. Though this technique shows good despeckling performance, it has no clearly defined cost function. This difficulty is addressed in the second contribution by utilizing regularization by denoising (RED), a framework closely associated with PnP. It defines an explicit prior using denoisers and has simple gradient expression. The data model is derived in the homomorphic domain. The cost function is devised by combining the data model with the RED prior and is solved using ADMM. Although the proposed technique increases despeckling accuracy, it has certain limitations. It imposes several constraints on denoisers, hence the benefits of the majority of denoisers cannot be realized. The third contribution is based on the actual priors that overcomes the problems associated with denoiser-induced priors. Specifically, the fractional order total variation (FrTV) Abstract prior and a non-convex penalty function are employed to preserve image features, avoid artifacts, and promote sparsity. The multiplicative noise model is converted to an additive model using suitable modifications. The cost function is constructed by combining the data term for the additive model with the aforementioned prior and penalty in the wavelet domain. The resulting optimization problem is solved using the well-known ADMM algorithm. It exhibits improved despeckling performance, but generates strip artifacts in practical SAR images. It is always a difficult task to choose an appropriate prior for regularized inversion problems. The fourth contribution exploits the benefits of nonlocal self similarity (NSS) and uses the dictionary based sparse framework to suppress the speckle noise. Specifically, the proposed sparse representation framework addresses the despeckling task by formulating an optimization problem. It consists of the data model derived in the multiplicative noise domain, an ℓ1−norm prior (sparse coefficients), and a penalty term that is based on the difference between the extracted patch and its corresponding sparse representation. It is solved by alternatively updating the dictionary, sparse coefficients, and the latent image. The simulations are carried out for both simulated and practical SAR images. The proposed algorithms are compared with several methods by computing the reference, without-reference, and edge-preserving measures. The despeckled and ratio images are visually compared for subjective evaluation. The fixed-point convergence analysis is carried out for the PnP-based method. Likewise, the residual convergence analysis is conducted for the RED- and FrTV-based algorithms. The merits and drawbacks of each strategy are discussed thoroughly in this research.