Adaptive Aggregation Network for Face Hallucination
Jin Guo,Jun Chen,Zhen Han,Han Liu,Zhongyuan Wang,Ruimin Hu
DOI: https://doi.org/10.1007/978-3-030-00767-6_18
2018-01-01
Abstract:Face hallucination refers to obtaining a clean face image from a degraded ones. The degraded face is assumed to be related to the clean face through the forward imaging model that account for blurring, sampling and noise. In recent years, many methods have been proposed and improved well progress. These methods usually learn a regression function to reconstruct the entire picture. However, there are huge differences among the optimal learned regression functions in different regions. In other words, the learned regression function needs to process all regions, which makes it difficult to reconstruct a satisfactory picture. As a result, the reconstructed images in some regions are relatively smooth. In order to address the problem, we present a novel face hallucination framework, called Adaptive Aggregation Network (AAN), which uses the aggregation network to guide face hallucination. Our network contains two branches: aggregation branch and generator branch. Specifically, our aggregation branch can explore regression function from low-resolution (LR) to high-resolution (HR) images in different regions, and aggregate the regions by the similarity of the regression function. Then generator module can be used to make a specific hallucination on the selected regions to get a better reconstruction result. After evaluating on datasets, our model was proved to be above the state-of-the-art methods in terms of effectiveness and accuracy.