Super-Resolution Imaging Using a Focus Pixel Sensor

Modern camera sensors are equipped with a focus pixel, a special type of pixel that can collect separate light rays from the left (L) and right (R) directions. The phase difference between the corresponding L/R pixels is utilized to facilitate quick auto-focusing. In this study, we expand the usability of the special pixels to super-resolve an image. We design a neural net to best fuse multiple low-resolution focus pixel images with a normal image based on repetitive channel and spatial attention layer structures. Empirical results show that focus pixel images contribute to the creation of fine details by providing additional information to super-resolve an image, especially for textured areas and that the proposed neural net-based method enhances the state-of-the-art super-resolution methods that do not use focus pixels in quantitative and qualitative measures.