Abstract
While the human visual system is able to adapt to a wide range of ambient illumination levels, cameras often deliver over- and/or under-exposed pictures of consequently low quality. This is particularly true of low-cost CMOS-based mobile camera devices that pervade the market. Towards finding a way to remediate this problem, we study the characteristics of poorly-exposed image regions under a natural scene statistics model with a goal of creating a framework for detecting, localizing and correcting overand/ or under-exposed pictures. Poorly-exposed picture regions are detected and located by analyzing the distributions of bandpass, divisively normalized pictures under a natural scene statistics model. Poor exposure levels lead to characteristic changes of the empirical probability density functions (histograms) of the processed pictures. This can be used to trace potential images saturated by over- or under exposure. Once detected, it is possible to ameliorate these distortions. If a stack (sequence) of maps of the same scene is available taken at different exposure levels, then it is possible to correct poorly exposed regions by fusing the multiple images. Experiments on multi-exposure datasets demonstrate the effectiveness of such an approach which suggests its potential for real-time camera tuning and post-editing of multiply exposed images.
Meeting abstract presented at VSS 2017