改进BP神经网络在手写数字识别中的性能研究  

The Performance Realization of Some Improved BP Neural Networkin Handwritten Digit Recognition

作　　者：李望晨[1] 张池平[2] 

机构地区：[1]潍坊医学院公共课教学部数学教研室,261042 [2]哈尔滨工业大学数学系,哈尔滨150001

出　　处：《信息技术与信息化》2006年第5期104-106,共3页Information Technology and Informatization

摘　　要：BP神经网络有网络结构本身的改进,基于传统梯度下降算法的改进和基于数值优化算法的改进三个主要方面。针对经过大量预处理后的较规范手写体数字样本,提取一种13维的结构特征向量,采用多种改进方式进行训练和识别测试以比较不同算法的性能。The improvement of BP neural network has three main aspects, such as the improvement based on network structure itself, the improvement based on traditional gradient descent algorithm and the improvement based on numerical optimization algorithm. A kind of 13 -dimention structure feature vector is extracted respectively from numerous standard handwritten digit samples after being preprocessed as network inputs. We use several improved methods to train test networks respectively and compare performance of different algorithms.

关 键 词：BP神经网络改进 手写体数字识别 LM—BP算法 特征提取 性能比较 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]