A new method of trademark retrieval based on average moments of feature points is proposed in the paper. First

the trademark image is divided into several concentric cells

and some special feature points are located in each sub-image according to the image's reference circle and principal orientation. The relative positions of these feature points have the invariability with respect to translation

scaling and rotation. After that

the new idea of average moments of these feature points is proposed. This feature descriptor has good robustness and is not so sensitive to noise and the delicate changes in image's edges. The experiments show that this method has good stability and can give attention to image's local features and global features at the same time. Therefore

the retrieval results match human visual percept ionk well.