Super-Resolution Reconstruction Via Feature Extraction Combine With Dictionary Learning

Image super-resolution reconstruction is a technology that can capture original high-resolution images from degradable low-resolution images, which has an important position in the field of digital image processing technology. Image super-resolution reconstruction techniques do not need to change the existed physical device, but the use of digital signal processing technology to realize the improvement of image resolution, and to achieve the purpose of overcoming the hardware precision and the cost of hardware. At present, the technology of image super-resolution reconstruction is widely used in remote sensing monitoring, high definition digital television, medical imaging, public security and other fields.At present, the image super-resolution reconstruction mainly have three kinds of methods, like methods based on interpolation, reconstruction and learning. In recent years, with the continuous development of sparse representation and the theory of low-rank matrix decomposition, methods of image super-resolution reconstruction based on sparse coding have become a research hot topic. Based on that, this paper focuses on the key step named feature extraction algorithm in the image super-resolution reconstruction and the construction technology of learning dictionary in sparse representation, and mainly study the super-resolution reconstruction of a single frame image by using patch haar wavelet feature extraction combine with sparse coding, also study the super-resolution reconstruction of multi frame images by using low-rank fusion combine with sparse coding. The contents of this paper mainly include the following three aspects:First of all, this paper elaborates basic theories of three types of image super-resolution reconstruction methods, and digitally realizes a part of algorithms, and also gives the corresponding experimental results and analysis, such as, the nearest neighbor interpolation, the bilinear interpolation and the bicubic interpolation which all based on interpolation reconstruction method, the iterative back projection algorithm and the projection onto convex sets algorithm which all based on reconstruction method, the example-based algorithm and the neighbor embedding algorithm which all based on learning algorithm.Secondly, the study of this paper based on the image super-resolution reconstruction algorithms (including the selection of image library, image features patch extraction and the training scheme of dictionary, etc.) of sparse coding, and proposes a new reconstruction method, which combine patch haar wavelet feature extraction with sparse coding of image super-resolution representation, and gives the corresponding experimental results and comparative analysis with other methods.Finally, this paper proposes a new method combined low-rank fusion with sparse coding of image super-resolution reconstruction, which based on the study of image registration algorithm and the theory of low-rank matrix decomposition, and gives the corresponding experimental results and comparative analysis with other methods.All in all, this paper puts forward and digitally realizes methods including super-resolution reconstruction of a single frame image based on the patch haar wavelet feature extraction combine with sparse encoding, and super-resolution reconstruction of multi frame images by using low-rank fusion combine with sparse coding, which around the super-resolution reconstruction of images. Further pointed out that the super-resolution reconstruction of multi frame images not only can fuse the effective information between images, but also can achieve better quality of images after super-resolution reconstructed by using priori information which provided from learning dictionary.