Abstarct: Super Resolution (SR) has been an active research topic in image processing. Due to the limitations and cost of high quality imaging systems, in many situations an affordable imaging system is used, instead using signal processing techniques the quality of captured image is improved. These techniques are specifically referred as Super-Resolution (SR) reconstruction. In this dissertation, we focus on self-learning approach for super resolution. Self-learning super-resolution methods utilize only input image for constructing a collection of training low and high-resolution image data. The goal of these methods is to estimate missing high-resolution details using information obtained from relationship between low and high-resolution pyramids that are constructed by input image. The images in the high-resolution pyramid are constructed using two steps down-sampling of the input image. The image at each level of the high-resolution pyramid is then up-sampled to construct the image at upper level in the low resolution pyramid. The relationship between low and corresponding high-resolution patches is modeled using support vector regression (SVR). Sparse representation and first-second order gradients are used as feature of each patches. We use sparse representation and first-order and second order gradients as feature of image patches. Using preserving adjacent information, and information of previous reconstruction are more attention in this dissertation. By considering adjacent pixels in non- smooth regions, the details of the image are removed. So, this method is useful for smooth regions in each images. To solve this problem, we introduce two methods baxsed on smooth and non-smooth regions. Two algorithms are presented baxsed on edge and non-edge pixels, separately. Applying one SR method on different texture images will result blurring and smoothing image. For this purpose, in this dissertation, content-baxsed image super resolution is introduced. Unlike existing methods, the images in the low-resolution pyramid are segmented and then used for the process of super-resolution. Another way to content-baxsed SR is to use sparse representation, neighborhood graph and clustering algorithms. For this purpose, k-means is used to cluster the data points before creating the similarity graph and data duplicate removal. Finally, the proposed methods in this dissertation are compared with several SR approaches, qualitatively and quantitatively.