In this paper we present a new method for self-localization on wafers using geometric hashing. The proposed technique is robust to image changes induced by process variations, as opposed to the traditional, correlation based methods. Moreover, it eliminates the need in training on reference patterns. Two enhancements are introduced to the basic geometric hashing scheme improving its performance and reliability: using quadtree for efficient data access and optimal rehashing for Bayesian voting. The approach proved to be highly reliable when tested on real wafer images.