The aim of Super Resolution (SR) algorithms is to produce a High Resolution (HR) image with the help of one or more than one low resolution (LR) image(s) of some real scene. Existing SR algorithms are of 3 types – Interpolation based SR algorithms, reconstruction based and Example learning based SR algorithms. Example based image SR has been recognized as an effective way to produce a HR image using an external training set. However, the effectiveness of these example based SR methods highly depends upon the consistency between the training set and the LR image(s). Single image SR method by learning multi scale self-similarities from an LR image has been proposed for the reduction of the effect created by incompatible high frequency details that are present in the training set. This proposed method is based upon the observation that in the natural images, small patches tend to repeat themselves redundantly many times, both within the same scale and across different scales. And to synthesize the details that are missing, HR-LR patch pairs have to be established using the LR image input and its down sampled version in order to capture the similarities that are present across different scales. Then Neighbor Embedding (NE) algorithm is applied to estimate the relationship between HR and LR image pairs. To completely exploit the similarities across various scales of an LR image, we use the previous resultant images as training examples for subsequent reconstruction process and adopt a gradual magnification scheme to upscale the LR input to the desired size step by step. Further, in order to preserve the sharper edges, we apply non local means (NLM) method to learn the similarity within the same scale. This has also suppressed the aliasing artifacts which are not required in an image.