Super-resolution Algorithm Research And Application Based On Dictionary Learning

High-resolution images have been widely used in various applications, such as satellite imaging, biometrics identification, pattern recognition and medical diagnosis. On the one hand, due to the limit of CCD pixel size, high-resolution images are hardly obtained. On the other hand, high-resolution CCD camera which has large volume and quality is expensive to construct. To meet the demand of low-cost high-resolution images, a technique named super-resolution arouses widespread concern.Super-resolution focuses on the problem of reconstructing a high-resolution im-age from one or more given low-resolution images. Generally, the existing methods can be categorized as interpolation-based, reconstruction-based and learning-based. Interpolation-based methods are the most common methods to enhance the resolu-tion of the given simple-image. However, super-resolving with interpolation-based methods is tending to yield overly smooth images with jagged and ringing artifacts. Reconstruction-based super-resolution methods are used if multiple low-resolution im-ages with the same scene are available. It is assumed that sub-pixel motion exists among these low-resolution images. Based on this assumption, a high-resolution im-age can generated by fusing all these low-resolution images. However, the performance of reconstruction-based super-resolution methods are unsatisfied when low-resolution images with sub-pixel motion are less, or the magnification factor is too large.Recently, learning-based super-resolution methods have been extensively studied in super-resolution. Different from the methods mentioned above, a set of training co-occurrence priors between low-and high-resolution image patches are employed in the learning-based methods. In this type of the methods, coupled dictionary based super-resolution algorithm is popular recently for its excellent performance in super-resolution. However, the process of training coupled dictionaries cannot be perfectly connected with the process of reconstructing super-resolution image in theory.Therefore, a novel conjugate dictionary based super-resolution algorithm is pro-posed in this paper. Different from the coupled dictionary based super-resolution al-gorithm, the low-and high-resolution dictionaries are trained separately. The low-resolution dictionary and the sparse representation are learned firstly, then the high-resolution dictionary is obtained by employing an L2-Boosting algorithm. Extensive experiments validate that our algorithm can surpass the coupled dictionary based super-resolution algorithm in both visual perception and statistical performance.