Spiking学习算法研究及其在图像特征提取上的应用

发布时间：2018-07-31 18:12

【摘要】：Spiking神经网络作为新一代人工神经网络,其时间编码的计算优势使其在研究领域的影响力与日俱增。在视觉神经系统的模拟层面,建立恰当的计算模型以模拟视网膜神经元的图像特征提取方式,并采用高效的学习算法对信息进行处理,一直是Spiking神经网络研究领域具有前瞻性和实用性的研究方向。本文在分析Spiking神经网络基本模型的基础上,从人类和灵长类动物模式识别的认知研究中得到了灵感,着眼于Spiking神经网络的经典学习规则以及神经网络图像特征提取技术,研究了基于Spiking神经网络的学习算法和图像特征提取技术的应用问题,主要包括以下内容:(1)研究并提出了一种具有良好特征表示的图像特征提取方法,该方法根据Spiking神经网络的特点改进了相位延迟编码方法对图像特征进行了转化。相位延迟编码作为Spiking神经网络常用的特征表示方法,具有很强的特征表达能力和生物可行性。通过对生物神经系统的体系结构、编码方式和学习理论的研究,我们考虑了外部刺激及不应期等多种因素,特征被最终转化为脉冲序列。通过信息的转换,图像自身携带的信息得以保留,取得了很好的特征提取效果。(2)在对经典膜电压函数相关的Spiking神经网络学习算法进行分析的基础上,改进了一种膜电压驱动的监督学习算法。该算法以目标输出脉冲时间点为标准,将学习情况分为目标输出脉冲时间点和非目标输出脉冲时间点进行约束和筛选,以提升学习的效率。算法通过减少图像特征的维度,在算法效率上要明显优于同类经典算法,而在膜电压背景噪声和输入抖动噪声存在的情况下,其鲁棒性也具有相当的优势。(3)提出了一种基于Spiking神经网络的图像识别新模型,新模型具有识别效率高,仿生性能好,鲁棒性强等特点。模型对图像特征进行了高效提取,保留了图像中关键的边缘信息和纹理信息,使用了更高效的学习算法处理输入模式的训练问题。整个模型从认知神经学的角度入手,猜想并模拟了生物神经网络从视觉输入到认知判断的过程,将理论应用到图像模式识别的具体问题上,通过对生物和计算科学的理论补充,完成了基础计算模型的建立。
[Abstract]:As a new generation of artificial neural network, Spiking neural network has more and more influence in the field of research because of its computational advantage of time coding. At the level of visual nervous system simulation, an appropriate computational model is established to simulate the image feature extraction of retinal neurons, and an efficient learning algorithm is used to process the information. It has always been a prospective and practical research direction in the field of Spiking neural network research. Based on the analysis of the basic model of Spiking neural network, this paper draws inspiration from the cognitive research of human and primate pattern recognition, focusing on the classical learning rules of Spiking neural network and the feature extraction technology of neural network image. The application of learning algorithm and image feature extraction technology based on Spiking neural network is studied. The main contents are as follows: (1) an image feature extraction method with good feature representation is proposed. According to the characteristics of Spiking neural network, the phase delay coding method is improved to transform the image features. As a common feature representation method of Spiking neural network, phase delay coding has strong feature expression ability and biological feasibility. Through the study of the system structure, coding mode and learning theory of the biological nervous system, we considered the external stimulation and the refractory period, and the characteristics were transformed into pulse sequence. Through the transformation of information, the information carried by the image itself can be preserved, and a good feature extraction effect is obtained. (2) based on the analysis of the Spiking neural network learning algorithm related to the classical membrane voltage function, A supervised learning algorithm for membrane voltage drive is improved. The algorithm takes the target output pulse time point as the standard, and divides the learning situation into target output pulse time point and non-target output pulse time point for constraint and selection, in order to improve the learning efficiency. By reducing the dimension of image features, the algorithm is more efficient than other classical algorithms, and when the voltage background noise and the input jitter noise exist, Its robustness also has some advantages. (3) A new image recognition model based on Spiking neural network is proposed. The new model is characterized by high recognition efficiency, good bionic performance and strong robustness. The model extracts the image features efficiently, preserves the key edge information and texture information in the image, and uses a more efficient learning algorithm to deal with the training problem of input pattern. The whole model starts from the perspective of cognitive neurology, conjectures and simulates the process of biological neural network from visual input to cognitive judgment, applies the theory to the specific problems of image pattern recognition, and complements the theory of biology and computational science. The foundation calculation model is established.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【参考文献】