基于渐进式学习的神经网络端到端验证码识别-《计算机技术与发展》

文章信息/Info

Title:: End-to-end Verification Code Identification of Neural Network Based on Progressive Learning

Author(s):: LIU Da-rong; ZHANG Yuan-ping; TANG Mao-bin; LI Fu-fang; School of Computer and Education Software,Guangzhou University,Guangzhou 510006,China

Keywords:: progressive learning; convolution neural network; verification code; no segmentation; end to end

摘要:: 针对验证码经过弯曲变形,无法采用传统的字符分割方法进行检测的问题,在模仿人类的渐进式学习过程的基础上,提出利用卷积神经网络,优化手写识别MNIST的三层网络结构,无需预先对验证码进行分割,直接对验证码进行端到端的识别.利用不确定度回收训练图片,减少训练集数据量,并利用可视化工具提高其网络识别性能.经过55万次训练后,生成了检测模型并对测试集验证码进行了检测,验证码识别速度达到0.073秒/张,准确率达到86％.通过对比同一测试环境下的测试集,发现利用渐进式学习方法具有更高的建模效率和更好的识别准确率,并对识别错误的验证码进行分析.

Abstract:: For authentication code can’t be measured by traditional character segmentation method after bending deformation,on the basis of the progressive learning of imitating human,we put forward convolution neural network to optimize three layers network structure of handwritten recognition MNIST,without prior segmentation for authentication code,directly identifying in the end-to-end to authentica- tion code. The training images are recovered according to uncertainty to reduce the amount of training data,and the network recognition performance is improved with visual tools. After training by 550 000 times,the detection model is generated and the verification code from the test set is detected. The recognition speed of the verification code reaches 0.073 seconds per piece,and the accuracy is 86%. By contrast with the test set in the same test environment,it is found that the progressive learning method has higher modeling efficiency and better recognition accuracy. The wrong verification code is analyzed.