Abstarct: Since 1970 CCD and CMOS sensors have been vastly utilized to make digital images. Although these sensors are suitable for some applications but their resolution are not suitable for specific applications in present and future. An approach which attracted the attention of many researchers for improving image resolution is the use of signal processing techniques to increase the resolution of low resolution images. One of main features for this approach is, no need to use high resolution hardware which increases the equipment cost. The methods which use this approach are known as super-resolution methods. In this thesis the objective is to reconstruct a super-resolution version of facial images through learning procedure. One of proposed methods in this field is known as LPS-GIS. In this method a training set using low and high resolution image pairs is collected. Then for a low resolution input image a number of low resolution images in the databaxse similar to it are selected. Using the corresponding high resolution images of the selected images in the databaxse a super-resolution version of the input image is constructed. In LPS-GIS, PCA was used to reduce the features in similarity criterion. In this thesis we use HOG, SIFT and SURF methods for feature extraction and the extracted features are then used to select those images in the databaxse similar to the input image. This is used instead of PCA method. We evaluated the proposed method and compare it with the LPS-GIS method.