Research On Super-resolution Based On Sparse Representation For Human Face Images

Since the image is an intuitional information carrier, the requirement for its quality is increasing with the rise of social informatization level. Image super-resolution as the effective measure to improve the resolution has become an important research field. In addition, due to the fact that the human face image works as a key role in the video surveillance application, super-resolution for human face images has become an important research topic on its own.This paper analyzes the super-resolution algorithms based on sparse representation in detail. On this basis, according to the characteristics of human face images, this paper puts forward a local adaptive super-resolution algorithm and a global adaptive super-resolution algorithm respectively for human face images. The local adaptive super-resolution algorithm makes up the shortage of the super resolution algorithm based on sparse representation that the dictionary pair for each block lacks adaptability. In the proposed algorithm the matching dictionary pair is self-adaptively selected from the dictionary library for each local block. If matching fails, the similar block training subset to the target low-resolution one is self-adaptively selected for suitable dictionary pair training. Then the new dictionary pair is updated into the dictionary library and used for recovering the high-resolution block. Additionally the matching threshold and the similarity threshold are self-adaptively adjusted according to the target low-resolution image. In the global adaptive super-resolution algorithm, considering the diversity of face images,the most suitable image training subset is self-adaptively selected from the total training set including a large number of different categories of images according to the global features of the target low-resolution human face image. Then the dictionary pair trained by the selected subset is used to recover the high-resolution image. The simulation experiments show that the two proposed algorithms achieve good effect on image reconstruction.This paper designs a super-resolution application software for human face images, which realizes the above mentioned two algorithms. It is verified that the software possesses high practicability.