Face Super Resolution Through Face Semantic and Structural Prior

Kalanke N.S.; Tomar A.S.; Arya K.V.; Inga Z.M.C.; Rodriguez C. · Lecture Notes in Networks and Systems (2023) · DOI: 10.1007/978-981-99-1912-3_15

Resumen

Face super-resolution (SR) is one of the domains of image SR that aims to recover a high-resolution (HR) image from low-resolution (LR) input. Existing deep-learning-based methods have significantly improved the performance, but most of them failed to preserve the semantics and structure in the output face images. The facial semantic and structural information determines the human face’s characteristics that need to be preserved during the SR process. Therefore, using facial semantics and structural priors, this research proposes a two-stage face SR framework. The first stage upsamples the LR image and utilizes the facial semantics (attribute information) prior which helps to generate the intermediate image with useful attribute information. After that, the second stage takes the intermediate image as input and extracts the facial features. The heatmap estimation unit is applied to estimate the landmark-based heatmap and fuses with extracted features that assist the proposed framework in preserving the structure in output face images. Numerous experiments shows that when reconstructing HR images from LR inputs, the suggested framework outperforms the state-of-the-art methods.

Instituciones: Universidad Nacional Mayor de San Marcos