A method of binary trademark retrieval based on sub-block images features is proposed in the paper. First, in order to
avoid the rotation effect, the trademark image's principal orientation is computed and the image is rotated accordingly.
Second, by way of locating the image's object region, the translation effect is removed as well. Then, the image's object
region is hierarchically partitioned into a quad tree of multilevel sub-block images. After that, the images' features are
extracted under different level of divisions, so that the multi-level shape descriptors can be gained. The experiments
prove the features based on sub-block images have strong invariability with respect to translation, scaling and rotation.
Besides, the features compromise image's local features and global features at the same time, therefore, the retrieval
results fit human visual perfect well.