High frequency compensated face hallucination with total variation constraint

Face Hallucination, one of the most famous learning-based image super-resolution techniques for facial images, can reconstruct a high-resolution image using only one low-resolution image. However, some detailed high-frequency components of the reconstructed image cannot be recovered using this method. In addition, the available LR images are sometimes blurred because of object movement or hardware problems. In this study, we proposed a high-frequency compensated face hallucination method with total variation constraint for HR image recovery. The proposed method is divided into two main processes: 1) Deblurring the LR input with total variation constraint; 2) Super-resolution with the proposed high-frequency compensated face hallucination. Experimental results show that the highresolution images obtained by our proposed approach are much better than those obtained by conventional methods.