An Improved Face Image Super-resolution Method Based On Neighborhood Embedding

With the development of the information age,digital image plays a more and more important role in the transmission of information in people's daily life.High resolution images are widely used in a wide variety of different application areas by virtue of their good visual experience.However,the images obtained are often lower in resolution due to the constraints of the hardware system or the photographing environment.It is very difficult to improve the resolution of the image by the improved hardware,so now we try to use the image super-resolution algorithm to increase the details of the image.Super-resolution technique is a software method based on information processing theory,which has a relatively low cost compared with the hardware based methods,and has a wide application prospect in the fields of security monitoring,digital communication and pattern recognition.This paper first introduces the research background of image super-resolution technique,further presents the development of it,and then the existing super-resolution techniques are summarized.More specifically,the learning-based method is carried out in details.As a branch of image super-resolution,human face super-resolution has always been a hot topic in the field of image processing,and the neighborhood embedding algorithm is universally recognized as one of the most effective face super-resolution algorithms,what's more,the two corresponding improvement method is put forward in view of its existing problems in this paper.The method firstly proposed can reconstruct high-quality high-resolution images,which is suitable for the environment with high requirement for image details.And the second method is mainly aimed at reducing the complexity of the algorithm and improving the speed of operation,which is suitable for the environment with real-time demand.The neighborhood embedding-based method assumes that HR manifold(HRM)and low resolution manifold(LRM)have similar geometric structure,and try to estimate HR image using the structure of LRM.However,the manifold assumption does not always hold due to the one-to-many relationship between the HR and LR image.The quality of the reconstructed image will get worse especially when the degree of image degradation is large.Moreover,the quality of the reconstructed image largely depends on the similarity between the test image and training set.Specially,the reconstructed result may be distorted when some particular features appear in the test face but not in the training set.To solve these problems,this paper proposes an improved face SR method based on multiscale and non-local similarity via a smooth path.Firstly,the multiscale similarity and non-local self similarity of the LR image are used to reduce the so-called dependence in consideration of the symmetry property of frontal face image.Secondly,we propose a multilayer scheme in which the HR face is reconstructed layer by layer,and the intermediate faces including both the test image and training set in each layer will make contribution to the reconstruction of the next layer,which will largely reduce the effect of the inconsistency of HRM and LRM.Finally,experiment and analysis of the algorithm in database are conducted as well as compared with the existing methods to evaluate the performance.The superiority of this algorithm is proved by subjective contrast and objective comparison.And we also conduct experiments on the real-world images to further testify the effectiveness of the proposed method.Neighborhood embedding-based method is not suitable for real-time applications for its high time complexity,the algorithm needs to go through the whole image training set whenever searching for neighborhood image blocks.To solve this problem,this paper proposed to improve the selection of eigenvalues and the neighborhood searching method.Specially,the traditional eigenvalues such as luminance or gradient are replaced by the singular values of the block matrix.Meantime,the image training set is firstly clustered before the neighborhood search,and then we compute the Euclidean distance between the test block and every cluster center,and further search the remaining neighborhood blocks according to the clustering center.Considering the structure of the high and low resolution simultaneously,the process of searching neighborhood patches is performed in high resolution space,and the weight calculation process is performed in low resolution space.Experimental results show that the proposed algorithm can shorten the running time of the algorithm and guarantee the quality of the reconstructed image at the same time.