Image Super-Resolution Reconstruction Based On Sparse Representation

Image super-resolution reconstruction is reconstructing a high-resolution image according to a single or multiple low-resolution images. It has very important applications in the fields of security monitoring, satellite remote sensing, medical imaging and so on. In the recent years, image super-resolution reconstruction based on sparse representation has become a hot issue. The prior information can be imported by dictionary learning, which improves the quality of the reconstructed image. Since part of information will lost for the sparse representation of images and image super-resolution reconstruction by wavelet analysis can efficiently utilize all information of the low-resolution image, the combination of these two methods will not only compensate the lost information by the sparse representation, but also can better use prior information brought by the sparse representation.In this paper, we proposed a new method based on wavelet transform. The low-resolution image will be interpolated by the nearest neighbor interpolation in the first. Then wavelet transform is performed on the interpolated image to obtain the low frequency subband, Haar wavelet is selected such that the low frequency can contain all the image information. For the high frequency subbands, we perform the wavelet transform on the Bicubic interpolated image by using Haar wavelet such that more information can be obtain in the high frequency subbands. Then the estimated high frequency subbands can be obtained by fusion the high frequency subbands obtained by the second wavelet transform and stationary wavelet transform. Finally, high-resolution image can be reconstructed according to the low frequency subband and the estimated high frequency subbands.Moreover, in order to impose more prior information, we propose a method based on sparse representation and wavelet transform for image super-resolution reconstruction. The high-resolution image obtained via sparse representation is decomposed by wavelet transform to get the three high frequency subbands, The low frequency subband is obtained by the wavelet transform of the nearest interpolated image. Experimental results show that for the visual effects, PSNR and structure similarity, our reconstructed high resolution image achieves a better result.