As traditional description vector calculation method used in image registration is too complex, time consuming and taking up more memory, a novel dictionary based local feature description algorithm was proposed. The K-singular Value Decomposition( KSVD ) method was used to generate dictionary and the feature descriptor was obtained by comparing the similarity between feature point region in images and elements in the dictionary. By above, the description vector generation algorithm was simplified and a higher feature matching speed was obtained. The matching process could be carried out by using randomized KD(k-dimension)tree algorithm. Then , the Random Sample Consensus (RANSAC) was used to choose the correct matching pairs. Finally, the transform parameters were estimated by using the least square method and the space geometric transformation of two images to be registrated was obtained. Results from experiments show that the proposed method reduces the description vector storage space, speeds up the feature matching and implements the registration process in real time.