Research On Super-resolution Reconstruction For Single Image Self-learning

With the rapid development of computer and digital technology,digital image has become an important way of storing and transmitting visual information in the field of computer vision.However,due to the influence of objective conditions and other factors,the resulting images are often low in resolution and poor in image quality in the actual imaging process.Therefore,based on the existing imaging equipment and low quality images,the super-resolution reconstruction technology that improves the resolution of the image by software is generated.Moreover,this technology has been widely used in the fields of medical imaging,security testing,craft repair and so on.Currently,learning based super-resolution reconstruction technology is a research hotspot in the super-resolution image reconstruction.The thesis mainly focuses on the problem of sparse representation?SR?image features in single image super-resolution reconstruction,and proposes an enhanced super-resolution reconstruction method oriented to pyramid image training sets and an improved super-resolution reconstruction method for non-pyramid image training set with fixed image size.Methods take SR and collaborative representation?CR?image features and the establishment of SVR models for low resolution?LR?and high resolution?HR?images as research main clues,and finally solve the problem of single image super-resolution reconstruction.It mainly includes:1.The implementation principle of super-resolution reconstruction method based on SR or SVR theory in recent years is studied and analyzed,establishing the key to super-resolution reconstruction is image features representation and the mapping relationship between LR image space and HR image space.For the problem of ignoring the characteristic of image continuity and equaling pixel effect when traditional method sparsely represents LR image block coefficients,an enhanced super-resolution reconstruction method is proposed,using the weighted idea to re-express LR image blocks.It emphasizes the role of central pixel in the construction sparse coefficients of image blocks,and reduces the reconstruction mapping error of the LR image blocks to the corresponding pixels of HR image;2.Aiming at the limitation of the SR process for image relies on dictionaries of the large training samples,and the problem that ₁?-norm tends to ignore a large number of image features.In sparse representation of LR image features,the CR method based on?₂-norm is used to weakly sparse,and an improved super-resolution reconstruction method that can preserve more image features is proposed.Compared to the super-resolution reconstruction method using SR based on?₁-norm,the improved super-resolution reconstruction method takes more image features into account when establishing the regression mapping model of LR\HR images,which greatly improves the reconstructed image quality.A large number of simulation experiments verify the feasibility and effectiveness of the super-resolution reconstruction method proposed in this thesis.Compare with other traditional super-resolution reconstruction methods,they achieve better results on multiple assessment indexes of image quality,and provide new ideas and methods for further research on super-resolution reconstruction algorithm.