The invention discloses a new neighborhood embedding-based super-resolution method. The method comprises the steps of firstly, performing two-dimensional discrete cosine transform (2D-DCT) on image blocks; secondly endowing a low-frequency coefficient in 2D-DCT with a relatively large weight value, and endowing a high-frequency coefficient with a relatively small weight value; thirdly, multiplying the coefficients of 2D-DCT by the weight values of the coefficients, and extracting features in an image through 3D-DCT according to a multiplying result; and fourthly, storing the result and real high-resolution block information in a training set. Meanwhile, an index for weighing the similarity between the image blocks by utilizing the result is proposed. The similar image blocks in the training set are selected by utilizing the index in an online amplification process, and the high-resolution image blocks are reconstructed by utilizing a manifold learning method. In order to reduce calculation complexity, an improved 2-level k-mean method is adopted for classifying the image blocks in the training set. An experimental result shows that the performance of the method is higher than that of conventional neighborhood embedding method and bilinear interpolation method.