The quadtree method is usually used in fractal block coding, but its efficiency is very low. We propose a new scheme, merged quadtree partitioning (MQP), for efficient image compression. It is an improved quadtree method, and can merge the quadtree nodes on the same level and on the different levels to share one transformation. As the merged nodes describe irregular regions with boundaries approximating the image edges, which need only one transformation, the total number of transformations required by MQP is much less than by the quadtree method. Thus, we obtain a greater compression ratio. In addition, we speed up MQP method by using the relationship between the search of the merged range and that of its parent range, and obtain a shorter encoding time than with other partitioning schemes.