Group Sparsity Penalized Contrast Source Solution Method for 2-D Non-Linear Inverse Scattering

Abstract

A group sparsity penalized CSI in the wavelet domain is proposed to alleviate ill-posedness within the framework of a contrast-source inversion (CSI) method. It is then applied to the retrieval of a large inhomogeneous dielectric scatterer from time-harmonic single-frequency data. As dependency exists between wavelet coefficients at different scales, referred to as the parent-child relationship, it enables to yield the wavelet quad tree structure. Therefore, wavelet coefficients can be regarded not only as pixel-wise sparse, but also group-wise sparse. Focus is put on using the dual-tree complex wavelet transform (CWT) to properly achieve the sought-after group-wise sparse representation of the spatial distribution of the contrast. It provides a 2,1 norm which is added to the standard cost functional to enforce group sparsity onto the wavelet coefficients of the spatially-varying contrast. The replication strategy is combined with the proximal method in order to solve the overlapping group penalized problem. Simulations from synthetic data in different configurations with in particular different signal-to-noise ratios illustrate pros and cons of the proposed method. The approach is shown to overcome the standard CSI method in demanding situations. Comparisons with the discrete wavelet transform (DWT) as usually performed and the 1 norm confirm the advantage of the proposed methodology.

Unrolled Convolutional Neural Network for Full-Wave Inverse Scattering

“…Yet, such local optimization methods may stall in local minima unless reasonable initialization. Prior information or regularizations such as positivity and sparsity [8] are usually enforced onto the optimization to alleviate illposedness.…”