基于级连分组BP网络的高精度手写数字识别

High Accuracy Handwritten Digits Recognition System with Cascaded Grouped BP Net

本文提出了一种基于级连分组网的手写数字识别的新方法。这种方法根据将每次分类任务简单化的思想 ,将每个网络的任务简化以提高其辨别能力。整个系统分为两级 ,第一级进行粗分类 ,选取前两个后选字。第二级对两个后选字进行细分类。每个细分类网络完成区分两类特定模式的任务 ,由于每个子网络分类数目的减少导致识别精度的提高。使用我们自己构造的含 1 0万个字符的库进行测试 ,我们的系统达到了在拒识为 5 %以内时误识为 0 .0 6 7%。

A novel approach based on cascaded grouped BP network is proposed to classify handwritten digitals.According to the concept of simplify each classification task,the number of classes for each net to classify is reduced to improve the accuracy.The whole system is divided into two cascades.The first step,rough classification is to select the first two candidates.The second,fine classification,is to classify two specific kind of patterns.So the accuracy is improved because of the simplification of each classification task.Our system achieves the substitute rate of 0.05% with the rejection rate of 5% based on our own database with about 100000 characters.