Abstract: An approach using both global and local image features to retrieve trademark image is presented.The principal orientation of a shape is first determined and rotated to become horizontal rightward.Then the object region on the rotated image is extracted and hierarchically partitioned into a quadtree of multilevel sub-blocks.Finally,sub-block image features are extracted to measure the image similarity.Experiments performed on a database containing 2000 images show that the sub-block image features keep good invariance under rotation,translation,scale and deformation,and the retrieved results match human visual perception very well.