Abstract:

In the processing of hallucinating faces reconstruction based on the sparse representation, aiming at the low sparse coding efficiency and precision caused by the redundant and over complete dictionary, a method is proposed, in which the clustered compact sub-dictionary is used to represent different objects of face images. The high resolution/low resolution (HR/LR) example face image patches are clustered by K-means algorithm. To make the compact sub-dictionary characterize the principal components of face image patches, the principal component analysis (PCA) algorithm is applied to learn sub-dictionary for each clustered dataset. After adaptively selecting the fitted sub-dictionary for a given LR face image patch, a centralized sparse constraint is added to enforce the sparse coding coefficients to approximate the HR face image patch to be reconstructed. The HR face image patch can be reconstructed from the line combination of the sparse coding coefficients and HR sub-dictionary. The final face image is synthesized by the HR patch and the smooth HR face image. With experiments on different face images and compared with other methods, the results demonstrate that the proposed method can hallucinate high quality faces in terms of both objective evaluation criteria and visual perception.