Small-World Architecture Based Kernel Auto-associative Memory Framework And Its Applications

This thesis mainly focuses on a small-world architecture based kernel auto-associative memory framework and its applications on small scale face recognition, specifically including how to construct a unified framework of conventional auto-associative memory models, how to reduce complexity of fully-connected models and the unified framework's application on face recognition. Some recent research theories, such as the popular "kernel method" in machine learning and extensively existed "small world network" in social fields, are applied to auto-associative memory models. Main contributions of this thesis are summarized as follows: Firstly, a unified framework of kernel auto-associative memory models is constructed and developed. In this thesis, based on Hopfield associative memory model, the author proposes a framework of kernel auto-association memory models by using the kernel trick to directly modify its recall (update) rules. With such a framework, we can not only uniformly formulate the existing auto-associative memory models but also derive more new different models using different kernels from the framework. Secondly, the complexity of fully-connected kernel auto-associative memory models is reduced. At present, almost all the auto-associative memory models adopt full-connected structure or architecture. With the real demands and the enlarging of the scale of handled problem, the complexity of the models have increased greatly, which will undoubtedly result in the difficulty of connecting up on the craft, and restrict models' circuit implementation. The research results in this thesis show that, compared with the regular network and random network, introducing small-world architecture into kernel auto-associative memory models is confirmed to be more effective. Finally, a series of novel face recognition algorithms based on associative memory models are proposed. On the basis of aforesaid work, the author further proposes robust face recognition algorithms based on sparse kernel auto- associative memory models. Simulation results on a part of the FERET face image database show that these algorithms not only achieve almost the same recognition performance as that of fully-connected models, but also outperform both famous Eigenfaces and (PC)2A algorithms after adding random noise or occluding partially on face image.