An image super-resolution (SR) reconstruction algorithm based on residual error is proposed. Patch pairs

composed of features for low-resolution (LR) patches and residual errors between original high-resolution (HR) image patches and interpolated LR image patches

are classified by K-means

Each class patch pair is trained by KSVD (K-singular value decomposition) to obtain an LR and HR dictionary pair. Residual errors are reconstructed by the dictionary pairs selected by the Euclidean distance between the test patches and class centers and by the weighted sum of the reconstructed results of the similar class patches. Then

combined with interpolated LR images and reconstructed residual errors

HR images are reconstructed. Experimental results show that the proposed method has a better performance and the method to classify patches and perform weight sum of the reconstructed results of the similar class patches is improving the quality of the SR image.