Selfie image super-resolution using an implicit prior learned from self-examples

Selfie image is a self-captured photograph of oneself using the front camera of a smartphone. The use of selfie images in social media is enormous and growing day by day. Most of the modern day smartphones are equipped with two cameras, viz. a high-resolution primary camera in the rear and a low-resolution secondary camera in the front of the smartphone. Typically selfie images are captured using low-resolution front camera and hence the spatial resolution of these images will be very much limited. In this paper, we propose an efficient approach to improve the spatial resolution of selfie images by exploiting the self-similarity across various scales of selfie images using a self-example based super-resolution algorithm. The super-resolution algorithm is formulated by learning a local regression from in-place self-example patches extracted from various scales of the given selfie image. In-place matching ensures that image patches extracted from different scales are spatially close and hence will have the same high-frequency details. A local regression is learned by approximating Taylors series and it serves as an efficient implicit prior to learn the relation between low-resolution and its corresponding high-resolution patch. The algorithm is evaluated both qualitatively and quantitatively and the results validates its efficiency.