Fast Exposure Fusion Based on Regression Analysis

Abstract

Facing a high dynamic range scene, cameras can not capture highlights and shadows in one single image. The main approach to deal with this problem is to capture several images with different exposure parameters, and then merge them into one image containing all details. The fusion process is performed as a weighted average between the corresponding pixels, where weights are assigned using quality measures on every pixel. The challenge is to produce image fusion with a small amount of operations to meet realtime requirements. In this paper we introduce a new exposure fusion technique based on regression analysis, which avoids the pixels weighting phase. Starting from a set of observations arising from an exposure fusion method, we estimate a regression model that imitates the fusion process. The aim of this technique is to avoid the considerable amount of computations required by the original exposure fusion method and then meet realtime fusion. Subjective and objective experimentations are shown to prove the effectiveness of the proposed technique.