Research For Superresolution Algorithms Based On Two-stage Dictionaries And Multi-frequency-band Dictionaries

As an active field in image processing, super-resolution has been widely appliedinto computer vision, medical imaging, image remote sensing and security detection,which aims to reconstruct high resolution image from one or multiple low resolutionimage(s). With the wide application of sparse representation in image super-resolution,two-stage and multi-frequency-band dictionaries are proposed for single imagesuper-resolution problem in this paper.Firstly, two-stage dictionaries are explored into recovering as much detailinformation as possible. Considering that there are many repetitive structures in naturalimage, nonlocal self-similarity information is combined properly with iterativeback-projection to post-process the image, i.e nonlocal iterative back-projection tofurther improve the reconstruction quality.Secondly, in addition to exploiting two-stage dictionaries to recover detailinformation, multi-frequency-band dictionaries consisting of low frequency (LF)dictionaries, middle frequency (MF) dictionaries and high frequency (HF) dictionariesare jointly learned to predict middle and high frequency information from low frequencycomponent by the prediction relation between LF\MF\HF components. Simultaneously,nonlocal iterative back-projection is applied into post-processing the image.Experimental results demonstrate the effectiveness of the proposed algorithm.Thirdly, the magnification factor of the traditional super-resolution algorithms onsparse representation is determined by high resolution\low resolution dictionary pairs. Toovercome the drawback, multi-frequency-band dictionaries are proposed forsuper-resolution problem only exploiting the prediction relation between LF\MF\HFcomponents to reconstruct the image. Considering the LF components consistency of theinterpolator image and original image, with the sparse representation of MF\HF patchesas the regularization condition, MF\HF patches are reconstructed exploiting thecoefficients similarity of LF\MF\HF patches. Furthermore, block matching and3D(BM3D) filtering is incorporated into iterative back-projection to postprocess the image nonlocally. Compared with other popular learning-based algorithms, the proposedmethod keeps better balance in reconstruction quality and speed.Finally, multiscale multiple-frequency-band dictionaries are constructed for SRproblem. The images reconstructed by dictionaries are fused by adaptive weightedaverage algorithm. Compared with other state-of-the-art algorithms, the proposed methodhas higher reconstruction quality and speed.