Structured Face Hallucination

Abstract

The goal of face hallucination is to generate highresolution images with fidelity from low-resolution ones. In contrast to existing methods based on patch similarity or holistic constraints in the image space, we propose to exploit local image structures for face hallucination. Each face image is represented in terms of facial components, contours and smooth regions. The image structure is maintained via matching gradients in the reconstructed highresolution output. For facial components, we align input images to generate accurate exemplars and transfer the high-frequency details for preserving structural consistency. For contours, we learn statistical priors to generate salient structures in the high-resolution images. A patch matching method is utilized on the smooth regions where the image gradients are preserved. Experimental results demonstrate that the proposed algorithm generates hallucinated face images with favorable quality and adaptability.

Compressed face hallucination

“…Due to space constraints, we only describe how the maps are generated in the JPEG compressed domain in this section. More details can be found in [9].…”

“…As a result, the color information needs to be better restored rather than bilinear interpolation [9]. In this work, we reconstruct HR outputs in color channels by patch match [14] to reconstruct HR outputs.…”

“…Figure 1 shows the main steps to generate the gradient map for a HR image from a LR input. The gradient maps U c and U e are generated in a way similar to [9]. Due to space constraints, we only describe how the maps are generated in the JPEG compressed domain in this section.…”

“…It is a well-known ill-posed problem as pixel values in the HR space need to be recovered based on a LR image with limited intensity or color information. Numerous methods have been proposed to address the face hallucination problem [1,2,3,4,5,6,7,8,9]. In [1], exemplar HR patches are exploited to overcome the limitation of insufficient highfrequency details in images generated by super-resolution algorithms with linear constraints.…”

“…Although local linear subspaces are exploited [7] to reduce ghost effects, highfrequency details of reconstructed HR images are missing (as subspace methods can be considered as low-pass filters for denoising). The exemplar-based method [9] decomposes face into several parts according to facial structure, while it generates pleasant visual quality for non-compressed images, it doesn't perform well for compressed inputs.…”