【作者】 李瑞；

【导师】 王军宁； 王虎元；

【作者基本信息】 西安电子科技大学 ， 电子与通信工程， 2015， 硕士

【摘要】 仿生神经网络经历了十几年的发展,学者们在模式识别的多个领域内对其展开了探索,但性能依然有待提高。本文首先介绍了仿生模式识别的知识和集成算法的知识,并提出了一种改进的仿生模式识别算法。本文的工作如下：首先通过对集成学习中AdaBoost算法的模仿设计了一种新的构造仿生神经网络的方法,该方法将每个仿生神经元都作为一个弱分类器,通过每次弱分类器的反馈来决定该仿生神经元在新构造的仿生神经网络中的权重。通过这种方法我们可以使构造出的仿生神经网络将注意力更集中于那些对结果影响较大的样本,从而能够提高传统仿生神经网络的分类能力；其次在构造单个仿生神经元时,考虑到图像空间像素点之间位置关系的影响,我们选用了比传统欧氏距离有更好表达能力的图像欧氏距离(IMED),这种对IMED的新应用能够增强单个仿生神经元的性能；最后在一些特定的环境下将IMED化简成一种卷积算子来降低计算要求,同时方便快速应用。最终通过本文算法构造的仿生神经网络的分类能力相对传统算法有一定的提升。最后我们在多种常用的数据集上验证了本文算法。结论：本文提出的改进的仿生神经网络算法经过实验的验证表现出比传统算法更好的性能。更多还原

【Abstract】 Bionic neural network experiences ten years of development. Researchers explore it in many fields of pattern recognition. but the performance still needs to be improved. This paper introduces the knowledge biomimetic pattern recognition algorithms and Boosting algorithm, then the author proposes an improved biomimetic pattern recognition algorithm. The work of this paper is as follow: firstly the author designs a new method of construction bionic neural networks by learning the AdaBoost algorithm of boost learning. In this method, every bionic neuron is looked as a weak classifier. Each weak classifier through feedback to determine the weight of this bionic neuron in the newly constructed bionic neural network. In this way we can construct a bionic neural network focuses more attention on the sample which has a larger impact of those results and improve the bionic neural network classification capabilities; secondly, in a single bionic neuron structure, take into account the influence of the positional relationship between the pixels of the image space, the author uses the Image Euclidean distance(IMED) which has a better performance than traditional Euclidean distance. This new application of IMED can enhance the performance of individual bionic neuron; thirdly the author simplifies the IMED as a convolution operator in some certain circumstances to reduce the amount of computation. The bionic neuron constructed by this operator improves the classification when it is used as a weak classifier.Finally, the algorithm is tested on a variety of commonly used data sets. Conclusion: The improved bionic neural network algorithm proposed by this paper shows better performance in proven experiments than the traditional algorithm.更多还原