DeepFuse: A Deep Unsupervised Approach for Exposure Fusion with Extreme\n Exposure Image Pairs

Abstract

We present a novel deep learning architecture for fusing static\nmulti-exposure images. Current multi-exposure fusion (MEF) approaches use\nhand-crafted features to fuse input sequence. However, the weak hand-crafted\nrepresentations are not robust to varying input conditions. Moreover, they\nperform poorly for extreme exposure image pairs. Thus, it is highly desirable\nto have a method that is robust to varying input conditions and capable of\nhandling extreme exposure without artifacts. Deep representations have known to\nbe robust to input conditions and have shown phenomenal performance in a\nsupervised setting. However, the stumbling block in using deep learning for MEF\nwas the lack of sufficient training data and an oracle to provide the\nground-truth for supervision. To address the above issues, we have gathered a\nlarge dataset of multi-exposure image stacks for training and to circumvent the\nneed for ground truth images, we propose an unsupervised deep learning\nframework for MEF utilizing a no-reference quality metric as loss function. The\nproposed approach uses a novel CNN architecture trained to learn the fusion\noperation without reference ground truth image. The model fuses a set of common\nlow level features extracted from each image to generate artifact-free\nperceptually pleasing results. We perform extensive quantitative and\nqualitative evaluation and show that the proposed technique outperforms\nexisting state-of-the-art approaches for a variety of natural images.\n