摘要
为获得具有强推广能力、高效的识别算法 ,提出了基于模糊超椭球聚类的模糊超椭球分类算法 ,并且用于无约束手写体数字的识别。模糊超椭球聚类能充分利用训练样本集的分布信息 ,运用较少的类别个数来表征复杂的样本分布 ,获得良好的识别效果和推广能力。在此基础上 ,模糊超椭球分类算法加入了有监督的控制 ,使算法在聚类过程中可以确定合适的类别数 ,使学习结果能更好地反映训练集的概率分布。然后 ,采用学习矢量量化等算法对其进行进一步有监督训练 ,从而取得更好的训练效果。在国际通用的 NIST字库和实际采集的手写体数字集进行的实验中 ,模糊超椭球分裂算法获得了令人满意的结果 ,而且具有进一步发展的潜力。
A self organizing network with the fuzzy hyperellipsoidal classifying (FHECF) algorithm was proposed to recognize handwritten numerals. The SOM clustering and the adaptive principle extraction (APEX) algorithm were used to reproduce the original learning result, with some small nodes including their coordinates and covariance matrices, to represent the main distributions of the training set. Then, the nodes that give worse performances are splited by fuzzy hyperellipsoidal clustering (FHEC) algorithm and the new nodes are modified to gain a better learning result. The algorithm identifies the suitable number of network nodes and the hyperellipsoidal classifying result to provide a more precise training requirement. With the help of supervised learning algorithms such as learning vector quantization (LVQ), the network gains a better performance. In experiments recognizing unconstrained handwritten numerals, the algorithm has satisfying performance.
出处
《清华大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2000年第9期120-124,共5页
Journal of Tsinghua University(Science and Technology)
基金
国家自然科学基金项目! (6 9775 0 0 1)
关键词
模糊超椭球分类
手写体字符识别
手写体数字识别
hyperellipsoidal clustering (HEC)
fuzzy hyperellipsoidal classifying (FHECF)
handwritten character recognition
learning vector quantization (LVQ)
