Image Super-Resolution (SR) algorithms based on sparse reconstruction generally require external training samples. The shortcoming of these algorithms is that the reconstruction quality depends on the similarity between the image to be reconstructed and the training sample. In order to solve the problem, an image super-resolution reconstruction algorithm based on local regression model was proposed. Using the fact that the local image structure would repeat in the corresponding position of different image scales, a first-order approximation model of the nonlinear mapping function from low to high resolution image patches was built for super-resolution reconstruction. The prior model of the nonlinear mapping function was established by handling the in-place example pair of the input image and its low frequency band image with dictionary learning. During the reconstruction of the image block, the non-local self-similarity of image was used and the first-order regression model was applied to multiple non-local self-similarity patches respectively, the high-resolution image patch could be obtained through weighted summing. The experimental results show that, compared with other super-resolution algorithms which also make use of image self-similarity, the average Peak Signal-to-Noise Ratio (PSNR) of the reconstructed images of the proposed algorithm is increased by 0.3~1.1 dB, and the subjective reconstruction effect of the proposed algorithm is improved significantly as well.