Multi-mechanism Neighbor Embedding Super Resolution Based On Compressive Self-coding Dictionary Representation

Super Resolution(SR) is an image quality improving method based on signal processing. Being a very valuable field in both research and applications, SR technique could provide a low cost solution for image quality enhancement in hardware resolution limited situations. This thesis aims at detail-preserved reconstruction and restoring dictionary coded signals with modified Neighbor-Embedding(NE) SR methods. Contributions of three aspects are presented as follows, including data transformation, small signal protection and error compensation.1. A Dictionary Representation based Neighbor Embedding(DR-NE) is given for detail preservation. Signal distance is mainly decided by large amplitude ones. To avoid this, NE algorithm is individually performed on different scale structures. With image represented as atoms, the distance calculation between patches is converted to similarity measurement between feature atoms, emphasizing small signals in neighbor searching. Experiments shows that, with back projection in addition, DR-NE averagely improves PSNR 0.1792 dB than NE. And the time cost is reduced by 52%.2. Introducing a sub-image super resolution method based on directional dictionary decomposition and predetermined neighborhood embedding(DD-PNE). Different amplitude sub-images are extracted by combined directional dictionaries to prevent small sub-images overwhelmed by large ones. In the followed recovery phase, to reduce the cost of neighbor searching, sub-images are reconstructed via predetermined neighborhoods which are obtained from sub-libraries in decision trees. Experiments shows that DD-PNE improves PSNR 2.93 dB on average and only 55.7% time cost of the original NE.3. To compensate errors caused by model inaccuracy, three improvements are proposed: a). Dictionary representation error is fed back(RERFD) to the results which gets the model amended; b). A HR image supervised correction(HRSC) method is proposed to compensate model residual. Contrast to back projection, HRSC down-samples the input, and then projects it into LR space. The SR model error between projected HR and the input is then interpolated and directly compensated into SR results to get the model supervised. c) To improve the interpolation noise produced in RERFD and HRSC, a residual pyramid based secondary compensation(RPSC) method is presented where the HRSC is used again in residual’s interpolation. Experiments showed that the proposed RERFD, HRSC and RPSC method are almost all effective for improving SR reconstruction results in different degrees.