人工脑基于同源同类事物连通本性的模式识别新神经网络模型研究

Artificial neural network is a scientific concept for the investigation of cognition of human brain as well as a effective model to mimic,extend, and expand pattern recognition of human brain. In practice, a number of remarkable outcomes on neural network model using pattern recognition have already been available, and also the accurate recognition rate of some neural network models have met the practical requirement. However, some conventional pattern recognition models face the bottleneck that the accurate rate of pattern recognition for some particular recognition objects can be hardly improved. According to the congition of connectivity nature of objects within the same source and same category, in this project several issues with regard to mimicing,extending, and expanding the pattern recognition of connectivity nature of human brain within the same source and same category will be researched using an artificial brain information processing method, a creative intelligent computing method, and an improved function extremum computing method, proposed by applicants. In addition, some new neural network models using pattern recognition of artificial brain based on connectivity nature within the same source and same category, which consists of neutron with formulas, network topology with connecting way between neutrons and values of connecting weight, and network algorithms for structural studying that ensures the direct connecting of local objects with the same source and same category and working. . The expected results of this project would maybe be able to complement and further develop the congnition research of pattern recognition of human brain, and also break through the bottleneck of some conventional pattern recognition models that accurate recognition rate can be hardly improved.

人工神经网络是研究人脑认知的科学概念。也是有效的模拟、延伸和扩展人脑模式识别的模型。实际上，模式识别的神经网络模型研究已经取得许多很好的成果，一些模式识别的神经网络模型的正确识别率也达到实用要求。然而，对于一定识别对象，一些传统的模式识别模型遇到了模式识别正确识别率很难再提高的瓶颈，本项目通过对同源同类事物连通本性的认知，采用申请人提出的人工脑信息处理方法、创新的智能计算方法和改进的函数极值计算方法研究基于同源同类事物连通本性的人脑模式识别的模拟、延伸和扩展等问题，拟提出并深入研究几类人工脑基于同源同类事物连通本性的模式识别新神经网络模型{神经元（计算式）、网络拓扑结构（神经元连接方式，连接权值取值）、网络算法（保同源同类事物局部直接连通的结构学习算法、工作算法）}。. 项目的预期成果有望丰富和发展对人脑模式识别的认知研究，突破一些传统模式识别模型遇到的正确识别率很难再提高的瓶颈。

人工神经网络是研究人脑认知的科学概念，也是有效地模拟、延伸和扩展人脑模式识别的模型，多年来获得了国内外学者的广泛关注。本项目通过对同源同类事物连通本性的认知，采用申请人提出的人工脑信息处理方法、创新的智能计算方法和改进的函数极值计算方法研究基于同源同类事物连通本性的人脑模式识别的模拟、延伸和扩展等问题，突破了一些传统模式识别模型遇到的正确识别率很难再提高的瓶颈。项目设计构造了基于同源同类事物连通本性的仿生模式识别的两个关键技术方案、基于同源同类事物连通本性的认证识别神经网络、基于同源同类事物连通本性的模式分类SLAM模型、保同源同类事物局部直接连通的模式分类SLAM模型、基于同源同类事物连通本性的分块并行优先度排序神经网络模型、基于同源同类事物连通本性的分块并行优先度排序RBF神经网络模型、基于同源同类事物连通本性的分块并行优先度排序DBF神经网络模型、基于同源同类事物连通本性的分块并行优先度排序通用高阶神经网络模型、基于同源同类事物连通本性的分块并行优先度排序SVM模型等。确保模型分类时能将同源同类事物分为同一类，进一步提高了已有模式识别的正确识别率。此外，项目组还对子空间学习、图嵌入、稀疏表示以及四元小波变换理论进行了拓展研究，提出了分块最大间距准则和分块二维最大间距准则，模糊的嵌入差分投影技术，模糊二维局部图嵌入鉴别分析技术，二维最大嵌入差分特征抽取技术，鉴别的图规则非负矩阵分解特征提取技术。项目组将这些模型和技术应用于人脸识别，纸币识别以及图像降噪中，取得了较好的结果，具有较高理论和实际指导意义。