Super Resolution Detection Based On Sparse Representation Of Multi-Dictionary

The aim of single image super-resolution is to reconstruct a high-resolution image from a single low-resolution input. With the development of science and technology, more and more fields need to use super-resolution method to find the high-resolution image, such as HDTV, remote sensing. However, improve image resolution from the hardware will lead to excessively high costs, so the super-resolution algorithm is proposed. Many super-resolution methods based on interpolation or reconstruction model are simple and easy to implement, whereas the performance is limited, for example, bilinear interpolation, bicubic interpolation, Projections onto Convex Sets (POCS) and so on. So, learning based reconstruction algorithm is proposed, from which a model has a better performance is learned based on data sets.In this paper, we propose a novel super-resolution algorithm based on multi-dictionary of spares coding, which employs the property that image with different exposure intensity can capture different kind of information. Firstly, we randomly select a large number of images with different exposure intensity from the database, and then the K-Means algorithm is adopted to cluster these images and their corresponding low resolution images obtained by down-sampling. Secondly, through solve the optimization function, we can obtain multi high resolution dictionaries and a low resolution dictionary. Thirdly, different exposure intensity images corresponding to the input image can be reestablished based on the high resolution and low resolution dictionaries, which is then integrated into one high resolution by gradient fusion method. Finally, image global constraint is conducted by the iterative back projection algorithm and a high resolution image rich of details is output eventually.In the experiment, the parameters of the sparse representation model are determined by experiments, and the validity of the K-means and the global constraint is proved. Finally, compared with the classical algorithm, it is proved that this model can be used to reconstruct the image with rich details.