Support Vector Regression Based Single Image Super-resolution

Due to the limitations of physical imaging systems and imaging environments, such as optical blurring, motion blurring, under-sampling and noise, it is not easy to obtain a desirable high-resolution(HR) image or an image sequence. As an effective technique to produce high-quality images using a low-cost imaging system, image super-resolution(SR) reconstruction has attracted considerable attentions and shows great potential for many practical applications including computer vision, video surveillance, remote-sensing images and so on. As we known, learning-based methods which conclude regression-based and coding-based algorithms are effective for single image super-resolution. This thesis gives a deep research on regression-based methods and proposes two different super-resolution algorithms.For regression-based single image super-resolution problem, the key is to establish a mapping relationship between high-resolution and low-resolution(LR) image patches for obtaining a visually pleasing High quality image. Existing support vector regression(SVR) based image super-resolution methods always utilize single layer SVR model to reconstruct source image, which are incapable of restoring the details and reduce the reconstruction quality. Concerning these issues, we present a multi-layer SVR model to describe the relationship between the low resolution image patches and the corresponding high resolution ones. Besides, considering the diverse context in the image, we introduce pixel-wise classification to divide pixels into different classes, such as horizontal edges, vertical edges and smooth areas, which is more conductive to highlight the local characteristics of the image. Moreover, the input elements to each SVR model are weighted respectively according to their corresponding output pixel’s space positions in the HR image. Experimental results show that, compared with several other learning-based SR algorithms, our method gains high-quality performance in terms of both visual quality and computational cost.The above approaches solve the regression-based problem by dividing the SVR model into several single-output regression problem, which obviously ignores the circumstance that a pixel within an HR patch affects the other spatially adjacent pixels during the training process, and thus tends to generate serious ringing artifacts in resultant HR image and increase computational burden. To alleviate these problems, we propose to use multi-task least-squares support vector regression machine(MTLS-SVR) to model the inherent spatial relationship between the HR and LR patches, which can effectively preserve sharp edges in the reconstructed HR image. In addition, the fact that image patches appear several times within the same scale and across different scales is of great help to improve the quality of reconstructed HR images. Therefore, we introduce a non-local self-similarity prior in natural images to formulate as a regularization term to further enhance the SORM-based SR results. Extensive SR experiments on various images indicate that the proposed method can achieve more promising performance than other state-of-the-art SR methods.