Reconstructing a High Dynamic Range Image With a Deeply Unsupervised Fusion Model

Abstract

To well record a high-dynamic-range (HDR) natural scene, multi-exposure images fusion is an affordable and convenient option, which is a hotspot in the field of HDR imaging. In this paper, we propose a deep learning-based method to address multi-exposure images fusion issue. Multi-exposure images are fed into the proposed network to output the optimal fused weights, and the weights are automatically learned and optimized instead of conventional hand-craftly setting. Besides, our proposed model is trained in an unsupervised way without the corresponding well-exposed images, which makes it more suitable in practice. Moreover, a novel customized loss function is proposed to boost the performance. According to the experimental results on both the benchmark dataset and the image sequences captured by ourselves, it is demonstrated that the proposed method outperforms the state-of-the-art methods in terms of both objective metrics and visual quality.

Keywords

• Customized loss function
• high-dynamic-range scene
• multi-exposure images fusion
• unsupervised deep learning network