Neural Augmented Exposure Interpolation for Two Large-Exposure-Ratio Images

摘要

Brightness order reversal could happen among shadow regions in a bright image and high-light regions in a dark image if two large-exposure-ratio images are fused directly by using existing multi-scale exposure fusion (MEF) algorithms. This problem can be addressed effectively via exposure interpolation. In this paper, a novel exposure interpolation algorithm is introduced by combining model-based and data-driven approaches to form a neural augmented interpolation framework. An image with a medium-exposure is initially interpolated by using intensity mapping functions (IMFs), and then refined via a novel exposedness aware network (EA-Net). Experimental results indicate that the model-based approach is improved by the data-driven approach, and the data-driven approach is benefited from the model-based method for fast convergence speed and learning with few training samples. The explainability of both the new EA-Net and the proposed framework is improved via such a neural augmentation.

关键词