Mental Rotation by Neural Network

This paper proposes a new and compact method for object image retrieval fusing Dominant Colors (DCs) and embedded Markov chain concepts. This proposed method uses combined color-texture features which are characterized in terms of their spatial interaction or interrelationship properties, modeled by means of a set of embedded Markov chains, each associated with a major spatial direction. Specifically, DCs are extracted from the object image, which are encountered pixel-wise along a given direction to form an embedded Markov chain. Normalizing the resultant Markov chains over all specified directions, the corresponding stationary distribution is derived and served as Markov Feature-Vector (MFV). We then employ the chi square distance between the feature vectors in comparing similarity of images. The MFV involves spatial structure information of both within and between dominant color regions. Moreover, it keeps simplicity, compactness, efficiency, and robustness. We conduct experiments using a comprehensive set of images including deformable shapes. Experimental results show that the proposed method can retrieve an important number of correct images with very high accuracy while the mismatch ratio remains constant.