| Many face images are captured by live cameras and often with low resolutionsdue to the environment or equipment limitations. How to recover human faces hasbecome an important problem in face analysis and recognition. This problem wasaddressed in the pioneering work of Baker and Kanade[1]as face hallucination.Some current methods work quite well under certain circumstances, such asless-degradation face restoration. However, low-resolution face image may degradeseriously due to variations in real-world applications such as in video surveillance,in this case, the face restoration is challenging.In this paper, a multi-scale face hallucination method is proposed to producehigh-resolution (HR) face images from low-resolution (LR) ones according to thespecific face characteristics and priors based on frequency bands analysis. In thefirst scale, the middle-resolution (MR) images are generated based on a patch-basedlearning method in DCT domain. In this scale, the DC coefficients and ACcoefficients are estimated separately. In the second scale, a DCT up-sampling forlow frequency band restoration and a high frequency band restoration are combinedto generate the final high-resolution face images. Extensive experiments show thatthe proposed algorithm achieves significant improvement.In summary, the contributions of this paper are consisted of the following threeitems:(1) A multi-scale face hallucination method is proposed, which can benefit froma progressive restoration by utilizing more reliable information obtained from theprevious scale. The reason of its rationality is that the learning-based model is farfrom sufficient in case of serious degradation. Therefore, we can benefit from aprogressive strategy.(2) The specific face characteristics and priors are explored differently based onfrequency bands analysis. Since natural images are often flat and smooth, highfrequency components exist often around edges. Hence, most energy lies at lowfrequency coefficients. However, some subtle characteristics which represent highfrequency information have crucial impact on visible quality. Therefore, it’sreasonable to deal with low and high frequency bands separately.The evidence in theexperiments shows that significant performance can be achieved even with relativesimple process and low computational costs.(3) A novel learning-based method built in DCT domain is proposed to generateMRIs which reduces the redundancy of training set thus meaning a simplifiedclustering-based offline training scheme. Besides, the basic idea of this method can be further extended to deal with realapplications such as face recognition of general image and video. |
