Using images of noise to estimate image processing behavior for image quality evaluation

Noise is an extremely important image quality factor. Camera manufacturers go to great lengths to source sensors and develop algorithms to minimize it. Illustrations of its effects are familiar, but it is not well known that noise itself, which is not constant over an image, can be represented as an image. Noise varies over images for two reasons. (1) Noise voltage in raw images is predicted to be proportional to a constant plus the square root of the number of photons reaching each pixel. (2) The most commonly applied image processing in consumer cameras, bilateral filtering [1], sharpens regions of the image near contrasty features such as edges and smooths (applies lowpass filtering to reduce noise) the image elsewhere. Noise is normally measured in flat, uniformly-illuminated patches, where bilateral filter smoothing has its maximum effect, often at the expense of fine detail. Significant insight into the behavior of image processing can be gained by measuring the noise throughout the image, not just in flat patches. We describe a method for obtaining noise images, then illustrate an important application— observing texture loss— and compare noise images for JPEG and raw-converted images. The method, derived from the EMVA 1288 analysis of flat-field images, requires the acquisition of a large number of identical images. It is somewhat cumbersome when individual image files need to be saved, but it’s fast and convenient when direct image acquisition is available.