Super-resolution For Single-frame Image Based On Non-local Dictionary Learning

Super-resolution image reconstruction technique improve the resolution of the image from one or multiple low resolution image by means of signal processing techniques to get high-resolution image. It is widely used in the video field, public safety, remote sensing imaging, medical imaging, communications, etc.Obviously it plays a vital role to the development of society, science and technology.This paper first introduces the research background and significance of super-resolution reconstruction technology, application field, research status, and the method of super-resolution technology has carried on the detailed introduction. On this basis, this article in view of the single frame image super-resolution reconstruction technique, the main work is as follows:(1) Using local SVD dictionary to reconstruct high resolution image. Global dictionary generally always need a large number of samples training to get a dictionary, and use this dictionary for whole image sparse representation, on the one hand, because of the training sample number, it has been a long time to train a dictionary, on the other hand because of the complexity of the image information, a dictionary can't contain all the information, and thus lead to the lack of some information, local dictionary as partial adaptability is strong, so it can get better details. So we introduce a local dictionary based on singular value decomposition, and then we will use local SVD dictionary for high-resolution image reconstruction. And we introduce how to training a local SVD dictionary from the low resolution image to reconstruct high-resolution image in detail.The experimental results show that the use of local SVD dictionary at high resolution image reconstruction can restore more details than global dictionary.(2) We propose the single-frame super-resolution image reconstruction algorithm based on non-local dictionary learning. This paper analyzes the advantages and disadvantages of reconstruction-based and learning-based super-resolution algorithm. Reconstruction-based methods can get more clear edge, but not accurate enough to restore the texture detail. Methods of learning-based generally get richer textures and details, but the effect of reducing the edges of the poor. Based on reconstruction and combining with the method based on learning, common as regularization item for constraints, can be integrated the advantages of both. Research shows that the image have self-similarity in different scales, due to the non-local means can better reflect the self-similarity of image in the same scale, and use the dictionary training from a low resolution image sparse representation high resolution image reflects the self-similarity of image at different scales, so the learing and reconstruction framework use non-local means and sparse representation as a joint regularization item. On the basis of the original framework, this paper has carried on the improvement to it. In order to highlight different scales of self-similarity and get better details information, we introduce the local dictionary based on SVD and proposed single-frame super-resolution image reconstruction algorithm based on non-local dictionary learning. Experiments show that the algorithm is whether in visual or evaluation of objective performance indicators have shown good improvement compared to the previous algorithm, the proposed algorithm is more robust to noise.