Abstract: Super-resolution image method based on canonical correlation analysis is a super-resolution reconstruction method in linear subspace, but canonical correlation analysis methods do not take full advantage of label information of the training sample class. Therefore, this paper presents an improved image super-resolution reconstruction method, namely using canonical correlation analysis added label information to maximize correlation between high-resolution images and low-resolution images in projection space for taking full advantage of classification information of the training samples. When reconstructing in coherent subspace, the method of sparse representation are used to choose the number of neighbors to increase the flexibility of the model, and then, converting the high resolution image corresponding to the low-resolution image test step by step. Experimental results show better results both visually and on the value of PSNR and SSIM.