The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical Engineering University of Southern California

Technical Report USC-IPI-690

“Estimation and Detection of Images Degraded by Film-Grain Noise”

by Firouz Naderi

September 1976

Film-grain noise is a term describing the intrinsic noise produced by a photographic emulsion during the process of recording an image on film. Although film-grain noise has been recently considered within the field of image processing, the nature of the noise is still somewhat misunderstood.

One goal of this study has been to use the theoretical and experimental results on film characteristics obtained by photographic scientists in order to define film-grain noise within the context of estimation theory. A detailed model for the overall photographic imaging system is presented. There are linear blurring effects at the initial and the final segments of this model to account for various optical and chemical degradations. The middle segment of the model represents signal dependence effects of film-grain noise and includes a nonlinear noise term. The accuracy of this model is tested by simulating images according to it and comparing the results to images of similar targets that were actually recorded on film. These comparisons point out that the model is a good representation of the photographic imaging system.

The restoration of images degraded by film-grain noise is considered in two different contexts - estimation theory and detection theory. Under the topic of estimation, a discrete Wiener filter is developed which explicitly allows for the signal-dependence of the noise. The filter adaptively alters its characteristics based on nonstationary first order statistics of an image. This filter is shown to have an advantage over the conventional Wiener filter.

In the case of extremely low contrast images digitized by a very small aperture, film-grain noise is so severe that conventional statistical restoration techniques have little effect. For use in this situation a heuristic algorithm is developed which incorporates some of the vision properties of the human observer. Bayesian detection theory is used to justify the procedure and to provide some insight into its use. This algorithm also explicitly includes the signal dependence of the noise and has the capability of greatly outperforming the human observer in locating objects corrupted by very severe noise.

Experimental results for modeling, the adaptive estimation filter and the Bayesian detection algorithm are presented.

To download the report in PDF format click here: USC-IPI-690.pdf (7.0Mb)