Abstract:

In order to enhance the resolution of single degraded images, a method of super-resolution reconstruction is proposed via over-complete sparse representation. The core of the super-resolution problem is to construct over-complete dictionary pairs and represent sparsely signals with respect to associated dictionary. For reducing the complexity of building dictionary pairs in training phase, the low-resolution dictionary is learned from patches but the high-resolution counterpart is numerically calculated using known sparse coefficients. In testing stage, a sparse representation of the low-resolution image over its dictionary is solved with a regularized orthogonal matching pursuit algorithm; thereby super-resolution reconstruction is realized by jointing an optimizing high-resolution dictionary. Experimental results demonstrate that, compared with other similar techniques, the peak signal-to-noise ratio (PSNR) gain of super resolved images is 3.3 dB, and the improvement of structural similarity (SSIM) is 0.09. Especially, both training efficiency and testing speed of this proposed algorithm have dramatically improvement. This approach can be applied over the high-ratio super-resolution reconstruction of single-frame blurred images, hence the resolution level of given images is effectively improved.