Feature Constrained Multi-example Based Image Super-resolution

Example-based super-resolution algorithm predicts unknown high-resolution image information by the relationship model learnt from the known high- and low-resolution image pairs. This kind of algorithm can produce high-quality images, but relies on large extern image database. We propose a multi- example based image super-resolution method constrained by image features. First, our method initially high-resolves the low-resolution image by the proposed feature-constrained polynomial interpolation method. Second, we consider low-frequency versions of high- and low-resolution images as the example pair. Each patch in the high-resolution low-frequency image searches its similar patches from the low-resolution image by adaptive KNN search algorithm, and the regression model between similar patches are learnt. Finally, the learnt model is applied to low-resolution low-frequency image to complement high-resolution high-frequency information. Extensive experiments show that the proposed method produces high-quality high-resolution images with high PSNR and SSIM values.