Research On Face Super-resolutionReconstruction Based On Example Learning

Face super-resolution (FSR) image reconstruction is a typical inverse problem incomputer vision. Different from traditional super-resolution, FSR aims to single image, the solution to which requires effectiveprior information as a supplement.Take advantage of the correspondence between cooccurrenceimage pairs at high and low resolutions, example-based FSR has provided effective models and data assurance. However, there are still two problems:one is how to improve the mapping accuracy from low-resolution (LR) to high-resolution(HR); the other is how to realize the optimization of selecting example candidate.Main study contents focus on how to improve the effectiveness of example learning in face super-resolution.The PCA-based methods are widely used to deal with facial image SR, they still have some disadvantages:1) the holistic PC A model tends to yield result like the mean face, which is not similar to the ground truthface;2) PCA comprises negative coefficients, which is hard to interpret itsreconstruction results. In view of this, we propose a novel parts-based face hallucination methodusing non-negative matrix factorization (NMF). In the first phase, we calculate the coefficients of NMF models for HR andLR training faces in order to to obtain the required feature pairs.Then, CCA is used to enhance coherence of the NMF coefficients for both HRand LR training images. At last,we propose a new residue compensation algorithm based on a new statistical model called Hidimensional Coupled NMF(HCNMF), which enhances the quality of high-resolutionhallucinated images by using the low-resolution residue.Experiment results demonstrate that the proposed method can achieve better visual effects and more convinced PSNR and SSIM.FSR based on example learning provides the required priors in reconstruction, utilizing the similarity of visual elements between different resolution. However, there are still two problems in practice.First, the optimal HR candidate patch may not be searched so that the reconstruction quality would be restricted. Second, the size of example may affect the reconstruction quality. If the example patch is too small, then it can not contain enough texture information. If the example patch is too large, then tiny texture information will be lost. In view of this, we propose a new method based on classified adaptive sparse representation. First, the patches are classified according to the difference between their structure features, in order to improve the accuracy of dictionary reconstrucion and to reduce the complexity. Second, adaptive training of example size is implemented in the patches of same category. Experiment results demonstrate that the proposed method can achieve better reconstruction effects than other methods.Due to the high complex and non-linaear face image variance, the relationship between high and low resolution example pairs must be non-linear. However, a large of FSR algorithms assume that the linear mapping relationship exists in high and low example pairs, which must lead to errors. In view of this, we propose a new method based on kernel method, which improves the mapping accuracy by transforming complex image data to high dimensional feature space. First, kernel locality preserving projection (KLPP) is used to make the mapping relation linear. Second, in order to obtain the required HR face image details, degrading model is used to ensure that the reconstruction result is similar to the ground truth.Experiment results demonstrate that the proposed method can produce available HR face image, which preserves the consistance between the result and the input image and reduces the image distortion.Extensive experiments have been conducted to evaluate the above proposed methods. Compared with other methods, distinct performanceimprovement in terms of both objective and subjective image reconstruction qualitydemonstrates the effectiveness of our methods.