Verification Code Recognition Based On Active Learning And Convolutional Neural Network
Xingqi Chen · 2021
Currently CNN has been widely used in the field of computer vision, but its training process relies on a large number of labeled data sets, and the labeling of the data sets will consume a lot of costs, and active learning can select the unlabeled data set which contains more information. The data can be annotated by experts, thereby reducing cost consumption, so the combination of the two can be well applied in the field of computer vision. This article combines active learning with convolutional neural network, and adopts a verification code recognition scheme based on image segmentation. The author will use the verification code identification scheme of the complete verification code to identify the verification code consisting of four digits. In addition, the author will use three query functions of Random sampling, Least Confident, and Margin Sampling to study the applicability and effect of active learning in verification code recognition. Both schemes have been tested on the captcha data set. The experimental results show that in the verification code recognition problem, convolutional neural networks based on active learning can provide the same accuracy as traditional neural networks under large-scale labeled data.