Empirical Characterization of Camera Noise - eScholarship
Roberto Manduchi, Jeremy Baumgartner, Markus Hinsche · 2013
Empirical Characterization of Camera Noise Jeremy Baumgartner 1 , Markus Hinsche 2 , and Roberto Manduchi 1 University of California, Santa Cruz, CA {jbaumgar,manduchi}@ucsc.edu Hasso Plattner Institute, Potsdam, Germany [email protected] Abstract. Noise characterization is important for several image pro- cessing operations such as denoising, thresholding, and HDR. This con- tribution describes a simple procedure to estimate the noise at an image for a particular camera as a function of exposure parameters (shutter time, gain) and of the irradiance at the pixel. Results are presented for a Pointgrey Firefly camera and are compared with a standard theoretical model of noise variance. Although the general characteristic of the noise reflects what predicted by the theoretical model, a number of discrepan- cies are found that deserve further investigation. Introduction The quantitative estimation of image noise is critical for basic operations such as denoising [5], thresholding [6], and HDR [2]. The simplifying assumption of “uniformly distributed Gaussian noise” in images is well known to be incor- rect: for the same camera, the statistical characteristics of noise depend on the exposure parameters as well as on the irradiance received by the pixel under consideration. Theoretical noise models and procedures for noise parameter es- timation have been described by several authors [3, 7, 4, 8]. In general, previous work either assumes access to raw data from the sensor, or tries to “reverse engi- neer” the image signal processor (ISP) that performs operations such as gamma correction, gamut mapping, and white point calibration, in order to estimate the “true” irradiance at a pixel and the noise characteristics of the acquisition process. Published work ranges from methods that assume a well-controlled il- lumination and reflection surface [3], to approaches that attempt to estimate relevant parameters from a single image [5, 1]. In this work we take an inter- mediate stance: we assume that a number of pictures of a stationary backdrop are taken with a number of different exposure settings, but make no particular assumption about the illumination and reflectance characteristics of the scene, except that they should remain constant during data acquisition. This can be easily achieved in a standard lab environment. For each exposure setting, pix- els with similar mean value of brightness are pooled together for noise variance estimation. This procedure produces a characterization of camera noise as a function of the mean brightness value and of exposure parameters of interest (shutter time and gain). The results can be used to validate theoretical models