Super-resolution Reconstruction Based On Image Self-similarity And Compressive Sensing

High-resolution image means owning high pixel density and rich image details, which is of great signification for satellite remote sensing imaging, medical imaging, target location and other special fields. To improve the resolution of image, in theory, we can not only start from the hardware but the software algorithm. But, in fact, due to the limitation of hardware cost and the complexity of production, almost all of the researchers enhances the image from the aspect of software algorithms.Super-resolution reconstruction is a kind of technology that improves the image resolution from the software algorithms, which is based on digital signal processing, machine learning and other theories. It can improve the original image resolution well, under the condition of not replacing the image sensing hardware equipment. Thus, it becomes the current research focus and has a wide range of applications, especially in medical and defense security. Super-resolution reconstruction can be divided into two parts: reconstruction based on single image and multiple images. In this paper, we focus on super-resolution reconstruction based on single image. After deeply studying the superiority and disadvantages of the previous super-resolution reconstruction technologies, we combine the advantages of image self-similarity and compressive sensing theory, then make some improvements on the existing reconstruction methods. The main contents are as follows:(1) A novel super-resolution image reconstruction algorithm based on gradient magnitude classification and self-similarity is proposed. Without increasing the external training samples, we introduce the patch rotation strategy, which enriches the diversity of training samples. To balance the time complexity increased by using the patch rotation strategy, the average gradient magnitude is used to classify the training samples and FLANN(Fast Library for Approximate Nearest Neighbors) is employed to search the similar image patches fast. Finally, the self-similarity weighting reconstruction is performed using the unfixed patch size. Besides, the final reconstructed image is optimized by the local constraints and iterative back projection algorithm. In constrast with the traditional self-similarity reconstruction algorithms, our proposed algorithm displays a better visual effect,both visually and quantitatively.(2) A improved super-resolution image reconstruction algorithm based on self-similarity and compressive sensing is proposed. It regards the image reconstructed on gradient magnitude classification and self-similarity as the high-resolution training image during the reconstruction process of compressive sensing. Then, the final reconstructed high-resolution image is enhanced by the unsharp masking algorithm. Compared with the previous ccompressive sensing reconstruction that uses the interpolated version of low-resolution input image as training image, our proposed algorithm makes full use of the image self-similarity, which makes the training high-resolution dictionary better and improves the reconstruction precision. Our algorithm can make the reconstructed color image much sharper, even if the magnification factor is larger, the reconstructed high-resolution images looks also ideal and satisfactory.