Performance of sparse coding based on image super resolution reconstruction model is influenced by dictionary selection. A dictionary learning algorithm based on collaborative sparse representation is proposed. In training stage, training image patches are grouped into different clusters by applying K-means clustering algorithm. A series of high- and low- resolution dictionaries are trained over every clusters by collaborative sparse dictionary learning model which is based on constraint of simultaneously sparse. The complex mapping relationship between low-and high-resolution image patches is transformed into a simple linear mapping by using an L2-norm based sparse coding model, and a series of mapping matrices corresponding to each different clusters are obtained. In reconstruction stage, each input image patch is mapped to a high-resolution patch by a mapping matrix which is selected by searching out the cluster with largest similarity to input patch. Experimental results show that the proposed method achieves better reconstruction quality by improving the dictionary learning process.