High-Dimension Biomimetic Information Science And Its Application In Pattern Recognition

Pattern recognition becomes to the hot issues of artificial telligence research at home and abroad.The traditional method of pattern recognition encounters a bottleneck when it is used to deal with high-dimensional data set with low-dimensional geometry structure. Recognizing and utilizing the internal geometry distribution characteristics of high dimensional data, simulating the way of human’s recognition are the new directions to solve the problem of pattern recognition. The High-dimension biomimetic information geometric recognizes and analyzes the data distribution from the view of geometric, and attempts to solve practical problems in artificial intelligence and pattern recognition by simulating the human thinking in images.Through the research on High-dimension biomimetic information geometric in the application of pattern recognition, we have found the following problems that have not been solved, such as how to cover the high-dimensional data set and how to extra the’CORE" of the data set? In the case of relatively large number of samples how to improve the recognition efficiency, and how to construct semi-supervised classification?For solving the above problems, we try to find the answers by combining the High-dimension biomimetic information geometric with the manifold learning and machine learning. Our contribution includes the following works:We have studied on the Hyper Sausage Neuron model and network learning algorithm of biomimetic patter recognition, and proposed the Multi-degree of Freedom Neurons Model and network learning algorithm. Further more we compared the advantages and disadvantage of the two kinds of neuron model through the face recognition experiments. And comparing with the SVM we improved that the biomimetic pattern recognition algorithm based on the neurons coverage is more suitable for the recognition problems of the’small sample’ database.We have also proposed an algorithm of kernel near point Fisher discriminant analysis combining with the manifold learning and linear dimensionality reduction for reducing the error of the near sample points. The algorithm aims at maximizing the distance of the heterogeneous data points in the projection subspace. And its effectiveness has been improved by the face recognition experiments.Combining with the theory of manifold learning, we proposed the biomimetic recognition algorithm based on the manifold distance (including two algorithms: principal manifold based on local PCA and tangent distance) for the inadequacies of the traditional biomimetic pattern recognition algorithm. And the manifold distance algorithm is proved more suitable for recognition problem of the’large sample" database than the traditional biomimetic pattern recognition algorithm by experiments. The traditional biomimetic pattern recognition algorithm approximates the sample distribution manifold by the simplex neuron covering while the manifold distance algorithm looks for the sample distribution manifold by dataset projecting to the "principal manifold". They are the two algorithms to solve the same problem. And they both belong to the supervised pattern recognition because the types of the samples are needed to be known.In further research, we find that the machine learning can get some illumination in High-dimension biomimetic information geometric. We proposed the relative Similarity’Path-based Spectral clustering algorithms. And it improves the sensitivity of the spectral clustering algorithm on the scale parameter in the Gaussian function.We also studied the application of the High-dimension biomimetic information geometric method, and which has been used in the feature extraction of image retrieval and image registration. The prototype systems for image retrieval and image registration are given in this paper. And the effectiveness of feature extraction methods is verified by the experiments.