Local Image Histograms for Learning Exposure Fusion

Local image histograms provide valuable descriptors of the behavior of an image around a given pixel. Rather than bucketing intensity values for all the pixels in an image, local histograms are defined separately for each pixel and represent the distribution of nearby intensities. Several image filters, including the median, mean-shift, and bilateral, can be described and computed using these histograms. In the SIGGRAPH 2010 conference, a new technique was introduced for efficiently computing such local histograms at all points in an image [2]. While previous algorithms had time complexity proportional to the support of the histogram, the new approach always runs in O(1) time per pixel, taking advantage of efficient Gaussian blur operations to simplify the local histogram computation process. While [2] applies local image histograms to the reexpression of previously-known image filters, these histograms also provide interesting features that could be used to apply machine learning to computational photography. In this project, we propose the application of this technique to learning reasonable approaches to exposure fusion, in which images of a scene taken at different exposures are linearly combined at each pixel to generate a meaningful output using the entire dynamic range of the display device. Ideally, although there obviously is a strong correlation between the histograms and values of adjacent pixels in a photograph, the histograms provide sufficiently strong descriptions of local behavior that this dependence can be ignored. Thus, we use examples of successfullyfused images and their per-pixel histograms as training data to learn a function mapping a single pixel and its corresponding histogram in each of the different exposures to a single output. This dependence likely is nonlinear, since local histograms often exhibit complex or even bimodal distributions. For this reason, we apply the “Least-Squares SVM” kernelized regression technique, which implicitly makes use of high-order features [3]. In the particular setting of image processing with locallyweighted histograms, evaluation of the learned exposure fusion function can be made considerably more efficient by taking advantage of the fact that histograms change with relatively low frequency across most images; a key contribution of this project is a new algorithm for efficient evaluation that makes image-processing-byexample a more feasible task. This project represents one of the first data-driven techniques in computational photography introduced to the graphics community. Its timings compare favorably with comparable approaches and more standard image processing techniques, and the output exhibits relatively few undesirable artifacts.