[Paper] Visible-Light Image Synthesis from Infrared Images Using Texture Transfer

抄録

We propose a novel method for synthesizing a visible-light image from a near-infrared one taken from the same viewpoint. Generating a multi-channel color image from a single-channel one is generally an ill-posed problem. As texture and color cues, the method refers to visible-light color images of the same scene but taken from a different point. The proposed method focuses on the estimation of the luminance component rather than the chrominance one because chrominance is not sensitive for human vision, and relatively easier to be estimated than luminance. Our cost function to optimize the luminance component is designed with the assumption that the synthesized image should be globally similar to the input infrared image structure and locally similar to the reference image patterns. Our experimental results show that the proposed method produces an artificial visible-light image, whose color appearance is more natural than color conversion methods, and geometrically more accurate than texture transfer-based methods.