Image Super-resolution Restoration Algorithm Research Via Improved Classification Dictionary

In practice, the quality of the captured image could not be perfect usually due to the limitation of the image acquisition device, ill-posed acquisition and so on. The Super-resolution(SR) is proposed to solve this problem. With the research of compressive sending and sparse representation, two novel SR algorithms are proposed in this paper.An image super-resolution restoration algorithm via improved classification dictionary is proposed to overcome the problem that the traditional sparse coding framework is unable to generate high-quality reconstruction images for using a generic over-complete dictionary. Firstly, we perform clustering on the training samples and train classified dictionary. And then use a fast human face detection based on skin color features and Adaboost to locate facial areas of an image in the sample library and low-resolution image for human face dictionary training. Finally, adaptively chooses the optimal subdictionary by each input low-resolution patch to form the corresponding high-resolution patch and uses human face dictionary for face region reconstruction.The reconstructed image quality of traditional learning-based SR algorithms is limited to the poor sample. An image super-resolution method via local self-similarity is proposed to solve this problem. Firstly, we bind several similar patches found in a limited window into a group. And then use the groups to infer the high-resolution image based on an image capturing model. Finally, set the high-resolution image as the initial reconstruction of SR algorithm instead of the bicubic interpolation image used in the traditional SR algorithm.Experimental simulation and comparison results show that, compared with other SR algorithms, the proposed methods provide better visual effect and more details.