An algorithm, which turns out to be highly efficient and runs fewer time, is proposed to extract the features from edges of image, thus the multi-scale image fusion and mosaic can be carried out. We build an edge-smoothing pyramid and extract the stable features for image registration. By reusing the multi-scale representation, the registered images are fused, and the cost of mosaic is eliminated. The demo results indicate that this algorithm can greatly eliminate the false feature match, advance the precision of transformation between images to sub-pixel, and condense the computation cost of registration and consequent mosaic. Finally, an experimental analysis for the high precision is presented.