Methods for Blind Estimation of the Variance of Mixed Noise and Their Performance Analysis

Abstract

Modern imaging systems provide a huge amount of images nowadays. These images are of different original quality. Some of them are practically ready for exploitation, e.g., visual inspection, object recognition, etc. Other ones need to be pre-processed as, e.g., by filtering, edge detection, segmentation, compression (Pratt, 2007; Bovik, 2000; Al-Shaykh&Mersereau, 1998), etc. In the latter case, it is desirable to know noise type and characteristics (Pratt, 2007; Elad, 2010). Such information is exploited by modern methods and algorithms of image denoising (Elad, 2010; Sendur&Selesnick, 2002; Donoho, 1995; Mallat, 1998), edge detection (Pratt, 2007; Touzi, 2002) for setting proper thresholds that depend on noise statistics.In some practical situations, noise type and basic characteristics are known in advance. An example is radar imaging by synthetic aperture radar (SAR) with known number of looks and image forming mode (Oliver&Quegan, 2004). However, there are quite many practical situations where noise type and/or characteristics are not known in advance. Images acquired by digital cameras can serve as an example where noise properties are determined by camera settings, illumination conditions (Liu et al., 2008; Foi et al., 2007), etc. Then, noise characteristics are to be estimated for each particular image subject to further processing, for example, filtering or compression (Liu et al., 2008; Foi et al., 2007; Lukin et al., 2011). Similar situation holds for hyperspectral imaging where noise properties and signal-to-noise ratio (SNR) depend upon sub-band and they vary considerably in different component (subband) images (Curran&Dungan, 1989; Uss et al., 2011).Note that below we mainly focus on considering multichannel images where the general term “multichannel” relates to color, multiand hyperspectral imaging, dual and multipolarization radar imaging, multitemporal sensing, where multiple images of the same scene or terrain are obtained. While for one or a few images it is sometimes possible to carry out manual (interactive) image analysis for the determination of noise type and characteristics, it becomes impossible or too labour-consuming to perform such actions for multichannel data, especially if estimation is to be done on-board or under conditions....