Novel neighbor embedding super resolution method for compressed images

In this paper, we focus on the super resolution (SR) reconstruction for compressed images. Compressed images have low quality by suffering from quantization errors and compression artifacts. We propose a novel learning-based SR method that is based on the local linear embedding (LLE). A new feature vector using the discrete cosine transform coefficients of the norm luminance is proposed. It represents the character of the image patch in the transform domain. Compared with several existing feature selections, this new feature vector can preserve edges better and capture more details. It is more suitable for the compressed image. We also optimize the training set. Experimental results show that our method gets better performance both in objective and subject.