Sparse Features for Multi-Exposure Fusion

Abstract

High dynamic range (HDR) capture and display devices can be used to approximately mimic the human perception of gamut of colors and fine details. However, the relative high-cost of these devices may currently make them be not affordable for many consumers. Multi-exposure image fusion (MEF) offers a cost-effective software-based solution to this problem. By fusing low dynamic range (LDR) images with different exposure levels, MEF aims to create HDR-like images for LDR display devices, that are high in quality but low in cost. This study proposes a novel MEF weight-map extraction method using sparse signal representations and k-means clustering. A preprocessing stage extracts initial masks from over- and underexposed images to be used for weight map extraction and the proposed clustering model allows the overall algorithm to have good fusion performance regardless of the number of input images contained in the input exposure sequence. After a final multi-scale pyramidal fusion, the resulting HDR-like images show not only visually pleasing but also statistically significant results when compared to state-of-the-art methods in the literature.