CNN-Based Deghosting in High Dynamic Range Imaging

Abstract

This paper addresses the challenging problem of avoiding ghost artifacts in compositing a high dynamic range (HDR) image by multi-exposed images fusion. We propose a convolutional neural network (CNN)-based deghosting algorithm, which treats the estimation of motion maps as a labelling problem. The network input consists of multiple low dynamic range (LDR) images presenting motion-related scene and exposure differences, and the output is a set of binary motion maps where black pixels indicate inconsistency between reference and non-reference images. The network is trained using simulated data, and we propose a novel approach for generating realistic synthetic multiple LDR images and their corresponding ground truth motion maps. The trained CNN model is tested on a variety of dynamic indoor and outdoor scenes. It outperforms the state-of-the-art without requiring camera meta data and is more flexible to handle higher-resolution images under limited GPU memory compared to end-to-end networks for HDR imaging. Our source code, pretrained models, and data will be made publicly available to facilitate future research.