Research On Recognition Of Clicked Chinese Captchas Based On YOLO V2

Information security has always been a hot topic of concern.Captchas are widely used as a means to protect users' information security on website,at the same time,the development of captcha technology also promotes the development of captcha cracking technology.Recently,a new technology called selected Chinese captchas has appeared in China.It requires the user to correctly click the Chinese character in the order of the prompt characters to successfully pass the verification,while improving the user experience,it also increases the difficulty of cracking.There is currently no paper on how to crack such captchas.The evolution and cracking of captchas is a mutually reinforcing process.Only by constantly updating captchas technology can the information security of websites be guaranteed in time.In this paper,the recognition of this kind of Captchas has been studied,and good results have been achieved.As there are more than 3,000 commonly used Chinese characters,the training network model needs too much data set for end-to-end's location and recognition,and it is impossible to manually label each Chinese character.This paper proposes to divide the recognition of such captchas into location and classification,which has achieved good results.Due to the lack of labeled samples of such Captchas,this paper proposes a weak supervised learning method for training combined with YOLO object detection algorithm.YOLO algorithm has poor detection effect for small targets and poor generalization ability for different aspect ratios of the same kind of target.The Chinese Captchas belong to the category of small targets and have different aspect ratios.In this paper,the YOLO algorithm model is improved by using the anchor mechanism of Faster R-CNN and the idea of full convolution network.Improving the full connection layer of YOLO by using the idea of full convolution network to improve the detection speed of YOLO.A Chinese classifier is trained by a small number of annotated pictures.Candidate boxes are obtained from the image by Region recommendation algorithm,and roughly annotated by the classifier.Roughly labeled images will be used to initialize YOLO's convolution network layer,thereby improving the detection performance of the model.The experimental results show that the improved method can train a small number of labeled images combined with a large number of unlabeled images,and achieve a large number of labeled images training model effect.Through anchor mechanism,the YOLO target boxes extraction layer is improved to learn the prior knowledge of the target boxes,so as to learn more strong features and improve the detection accuracy of the model for small targets.Using Python language to generate common training images of 3755 Chinese characters,including 4 fonts and several image enhancement operations to ensure the generalization ability of the network.Then build a deep convolutional neural network to train these data sets which has achieved good results in the recognition of such captchas.In order to facilitate the display of the recognition effect,this paper finally realizes position of the corresponding Chinese characters in order after recognizing the Chinese characters through Python,thus giving an intuitive display effect.