优化卷积神经网络在复杂验证码图片识别中的应用  

Application of Optimized Convolutional Neural Network in Complex Captcha Image Recognition

作　　者：张爱民 ZHANG Aimin(Party School of Guoyang County Party Committee, Guoyang 233600, China)

机构地区：[1]中共涡阳县委党校,安徽涡阳233600

出　　处：《信息工程大学学报》2021年第6期670-674,共5页Journal of Information Engineering University

摘　　要：常规方法预处理验证码图片时,往往容易忽略边缘离散噪声点的去噪处理,造成预处理后图像峰值信噪比、结构相似度较低。为了解决上述问题,提出优化卷积神经网络在复杂验证码图片识别中的应用。转化原始图片为灰度图像,利用高斯平滑和二值图像逻辑运算,分别去噪正态分布噪声点和边缘离散噪声点,分割预处理后的图片,提取图片字符特征向量,统一标注后输入优化卷积神经网络,优化网络权值和变换参数,通过迭代训练,输出识别的图片字符信息。抽取复杂验证码图片,设置对比实验,结果表明设计方法相比常规方法,预处理后图像峰值信噪比分别提高了8.4 dB、13.5 dB、14.9 dB,结构相似度分别提高了3.8百分点、8.1百分点、12.8百分点,优化了图像去噪效果,降低了图片识别过程复杂程度。When preprocessing captcha images with conventional methods,it is easy to ignore the denoising of edge discrete noise points,resulting in low peak signal-to-noise ratio and structural similarity of preprocessed images.To solve the above problems,the optimized convolutional neural network is proposed to be applied in complex CAPTCHA image recognition.First,the original image is transformed to gray image,and Gaussian smoothing and binary image logic operations are used respectively to denoise normal distribution noise points and discrete noise points.Then,the pretreated image is segmented to extract the image character feature vector,which is input into the input optimization convolution neural network with unified mark.Last,through network weights optimization,parameters transforming and iterative training,the identified image character information is output.Complex CAPTCHA images are extracted,and a comparative experiment is set up.The results show that compared with the conventional method,the PSNR of preprocessed images increase by 8.4 dB,13.5 dB and 14.9 dB,respectively,and the structural similarity increase by 3.8 percentage point,8.1 percentage point and 12.8 percentage point,respectively.The image denoising effect is thus optimized,and the complexity of image recognition process is reduced.

关 键 词：验证码图片 字符信息 预处理 特征向量 神经网络 信息识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]