Super-resolution reconstruction of face images based on pre-amplification non-negative restricted neighborhood embedding

The traditional super-resolution (SR) reconstruction algorithm based on neighborhood embedding preserves the local geometric structure of image block manifold to reconstruct high-resolution (HR) manifold. However, when the magnification is large, the low resolution (LR) image is seriously degraded and most of the information is lost after down-sampling. The neighborhood relation of the LR manifold can not reflect the inherent data structure. In order to solve the problem effectively, we propose a face image SR algorithm based on pre-amplification non-negative restricted neighborhood embedding. In the training phase, the LR image is pre-amplified so that there are more similar manifold structures between the HR and LR resolution images. The constraints of the reconstructed coefficients are loosened and the HR image blocks are iteratively updated to obtain the reconstructed weights. The experimental results show that the proposed method has a better reconstruction effect compared with some traditional learning algorithms.