Multiple-Exposure Fusion with Halo-Free Convolutional Neural Network

Abstract

The dynamic range of the imaging device represents its ability to capture bright and dark targets in the scene. Limited by the hardware, the dynamic range of a single imaging will lead the loss of information like over-exposed or under-exposed, which makes the look and feel of the imaging result unsatisfactory. Although the dynamic range of imaging can be expanded through multi-exposure fusion, there is risk to produce artifacts such as halos. To address the above issue, an Anisotropic Convolutional Block based on convolutional neural networks is proposed, which can inhibit the halo among the edges with high contrast. At the same time, a fusion strategy based on image structure similarity and pixel intensity is proposed, which can improve the visual perception of imaging results. Experimental results prove that the proposed method can effectively improve the quality of high dynamic range imaging.