HDR Video Synthesis Using Superpixel-Based Illuminance Invariant Motion Estimation

Abstract

The dynamic range of real scenes is much bigger that what can be captured by consumer capturing devices. This constraint motivated the advancement of digital imaging technology to develop an alternative approach, i.e., high dynamic range (HDR) imaging. Conventional HDR image synthesis usually faces the problem of dynamic scenes while capturing the input images. This had led to the problem of ghosting artifacts, which has inspired the development of deghosting techniques. Besides the trend of HDR imaging, it is necessary to consider HDR videos. In some manner, HDR video can be considered as the extension of HDR imaging with dynamic objects. However, the synthesis of HDR video has more constraints, such as the number of input frames and the motions between adjacent frames. The most important problem with synthesizing an HDR video is how to estimate the motion map between two consecutive frames so that ghosting artifacts do not appear in the final HDR frame. Although several approaches have been developed, they still have a problem with the computational time or the problem of an inaccurate motion map due to the non-rigid motions. In this thesis, we propose a novel HDR video synthesis framework using illuminance-invariant superpixel-based motion estimation. The proposed algorithm first chooses an input image as the reference and employs a descriptor that is consistent with the illuminance variation and superpixel to esimate the motion between two adjacent frames. The input frames are then warped to the reference frame using the estimated motion maps. Finally, we construct a weight map to synthesize the final HDR frame. We perform the experiments on both real and synthetic data sets and compare the results with state-of-the-art algorithms. Experimental results show that the proposed algorithm provides HDR videos of higher quality than those of conventional algorithms under both subjective and objective assessments.