The intrinsic error of exposure fusion for HDR imaging, and a way to reduce it

Human visual system is capable to capture details in dark and bright areas at the same time despite the dynamic range of the scene. Digital cameras, however, have a limited dynamic range, therefore only capturing dark or bright areas in a single shot. Considering that, Mann and Picard [6] presented a method for the creation of high dynamic range (HDR) radiance map by exposure fusion of a stack of low dynamic range (LDR) images. There is a wide range of HDR literature: assuming static scenes and no camera motions [2, 6], considering motion on camera and/or in the scene [4], and most recently video creation [5]. Camera sensor provides the irradiance, RAW values (which are proportional to the radiance scene), but then cameras apply linear and nonlinear transformations, and a quantization step are applied resulting in a non-RAW image (such as JPEG). All approaches based on exposure fusion applied on these non-RAW images share a set of assumptions: 1) different color channels are independent, 2) in-camera non-linear correction curves are monotonic and smooth, but arbitrary in shape, and 3) the camera response function (CRF) remains constant while changing the exposure. Our work shows that these assumptions are not preserved when working with non-RAW formats, because camera may introduce changes from one exposure to the next. Thus, the exposure fusion of non-linear images will introduce more error in the final HDR radiance map, than computing it directly from RAW linear data.