Efficient High Dynamic Range Video Based On Convolutional Neural Network

High dynamic range(HDR)technology aims to solve the large gap of luminance dynamic range between real world and digital image systems.It has been widely used in areas like photographing,surveillance and photorealistic rendering.Currently,the main problem of HDR video is the presence of motion across frames and the need for temporal coherence,which make it hard to avoid ghosting artifacts.Specifically,many algorithms' performance and efficiency are often impacted by the unreliable process of aligning frames of different exposures.As most classic algorithms of image aligning and motion estimation like optical flow algorithm,are unable to handle such scenario.This thesis researches on the problems present in current HDR video algorithms.We focuses on the key step of reconstructing frames of missing exposure,while trying to solve the key problem of aligning frames of different exposures,and sum up an efficient new HDR video method based on CNN.Firstly,we present a new method that use CNN to do multi-exposure motion estimation.By training CNN that extract and match motion features under exposure change on a custom flow dataset of multiple exposures,we achieve a CNN model that can efficiently complete task of motion estimation between images of different exposures.In addition,using this as a base,we also develop an effective and concise framework to handle the whole HDR video reconstruction process with increased efficiency while maintaining good performance.A series of experiments were conducted to test,evaluate and compare our methods with other algorithms.The results demonstrate that the CNN based multi-exposure motion estimation method and the new HDR video reconstruction procedure both have good performance and relatively better efficiency.