Besta-Divergence-Based Variational Model for Speckle Reduction
Hyenkyun Woo, Junhong Ha · IEEE Signal Processing Letters · 2016
The image data captured through a synthetic aperture radar (SAR) system is in general corrupted by multiplicative speckle noise. Due to the heavy multiplicative structure of speckle, it is extremely difficult to identify objects in an SAR image without any denoising process. Therefore, in this letter, we introduce a new beta-divergence-based variational model with total variation (beta-TV). The main advantage of the proposed beta-TV model is that it reveals a natural connection between the maximum a posteriori (MAP) estimation based nonconvex variational model and the I-divergence-based convex model without any special transform, such as log-transform or m th root transform. Through this relation, we show that beta-TV can be used as a convex-like relaxation of the MAP-based nonconvex variational model. Empirically, we observed that the beta-TV model obtains a better peak signal to noise ratio (PSNR) compared to the state-of-the-art models, such as I-divergence-based model and the m th root transform based model.