小数据集复杂文本验证码识别  

Complex text CAPTCHA recognition with small data set

作　　者：陈诗雨 CHEN Shiyu(Wuhan Institute of Posts and Telecommunications,Wuhan 430070,China)

机构地区：[1]武汉邮电科学研究院,湖北武汉430070

出　　处：《电子设计工程》2024年第3期54-58,共5页Electronic Design Engineering

摘　　要：针对小数据集复杂文本验证码识别,使用传统的机器学习方法存在去噪难和分割难的问题,使用深度学习方法需要大量数据增加泛化性,数据不足容易导致过拟合。根据验证码的特点,提出了基于字符分割的小数据集复杂文本验证码识别方法,使用pix2pix做文本去噪,使用目标检测网络YOLOv4分割单字符,使用KNN算法进行分类。实验结果表明,该方法的单字符识别准确率达98.7%,优于开源模型的识别结果,验证了该方法在破解小数据集复杂文本验证码上的可行性和有效性。For the problem of complex text CAPTCHA recognition with small data set,using traditional machine learning methods is difficult to segment and denoise,using deep learning methods requires a amount of data to increase generalization,otherwise insufficient data is easy to lead to over⁃fitting.According to the characteristics of the CAPTCHA,this paper proposes a character segmentation based recognition method for complex text CAPTCHA in small data set.Use pix2pix to remove noise.Use YOLOv4 to segment single character.KNN is used for classification.The result of the experiment show that the single character recognition accuracy of the proposed method is 98.7%,it is better than the recognition result of the open source model,which verifies feasibility and effectiveness of the method.

关 键 词：验证码 深度学习 目标检测 KNN 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]