Image And Video Signal Super-resolution Reconstruction Based On Non-local Similarity

Due to the limitations of physical resolution of the terminal devices or the bandwidth in the transmission process,it is difficult to get the high-resolution images that meet the basic requirement of applications.To overcome this difficulty,super-resolution is proposed to reconstruct a high-resolution image based on the existing low-resolution images,which has been widely used in many fields,such as satellite remote sensing,digital entertainment,video surveillance and so on.Image spatial resolution has been improved by digging deeply into the information embedded in image,however,many super-resolution algorithms ignore the non-local similarity of image,resulting in the waste of priori information,there still exists room to improve the performance.To resolve the problems above,this thesis mainly focuses on image super-resolution based on compressed sensing and similarity constraint,and multi-frame video super-resolution based on non-local similarity.For image super-resolution algorithm,an image super-resolution based on compressed sensing and similarity constraint is proposed in this thesis: 1)A dictionary classification method based on measurement domain is proposed,image blocks are divided into smooth,texture and edge parts,the corresponding dictionaries have been trained using the classified image blocks,which can improve the representation of dictionaries.2)In reconstruction,to solve the problem that learning prior knowledge only from external training sets causing the false detail result,the image non-local similarity is fully used,all similarity image blocks are searched in the whole image,and the sparsity and similarity of blocks are both took as constraints in reconstruction process to improve the authenticity of the high resolution image.For video super-resolution algorithm,a multi-frame video super-resolution based on non-local similarity is proposed in this thesis: 1)A hierarchical optical flow based on adaptive smoothing strategy is proposed to estimate motion vector,including image preprocessing to reduce the effects of external lighting,adaptive smoothing strategy in flat area to improve estimation accuracy,warp each layer image using result vector of upper layer to solve the problem of poor accuracy in large displacement;2)The multi-frameimages fusion is realized by MAP theory,and a non-local model based on gradient edge enhancement is presented,for the regularization term,the similar weight of two pixels is calculated by image non-local similarity,and the gradient operator is introduced to enhance the effect of the edge part,to ensure the reconstructed image has a good edge effect.Finally,the experimental results show that the proposed super-resolution algorithms in this thesis have good performance in both subjective and objective aspects,and can be applied to practical scenes such as mobile entertainment and video surveillance.