Statistical modeling of image noise and its applications in computer vision

Abstract

Noise impacts practically all aspects of computer vision, especially for computing image derivative and removing noise. Unreliable estimation of image derivative and noise removal caused by inherent noise characteristics makes subsequent applications, such as feature detection and photometric approaches, produce inconsistent outputs. In this thesis, we develop a novel noise model for computer vision applications related with image derivatives and noise removal. In the first part of the thesis, we present a sensor noise model, which is based on physical models of image noise in the image formation process. We show that the widely used additive Gaussian noise model is inadequate under general illumination conditions. We present an intensity difference based model to overcome the limitation of the previous model. Using the dominant photon noise assumption, image noise for intensity difference can be a distribution model represented by the Skellam distribution, derived from the Poisson distribution of photons. We show that intensity differences caused by noise fits the Skellam distribution well in the spatial domain, as well as in the temporal domain. Our modeling shows the linear relationship between intensity and the corresponding noise parameters of the Skellam distribution even under natural illumination conditions, while conventional variance computation of the Gaussian distribution does not. Because the Skellam parameter obtained from the linearity determines the distribution of noise for each intensity, we can statistically measure intensity difference using the estimated distribution of noise. In the second part of the thesis, based on our noise model, we reconsider fundamental applications related with image derivatives and noise removal. First, given a confidence coefficient, we can determine an intensity range corresponding to variation caused by image noise for each pixel using the estimated noise distribution. We apply the intensity range to e...