CAPTCHA Recognition with Adaptable Neural Networks
- CAPTCHAs are tests to tell computers and humans apart. Text-based CAPTCHAs are very common although there exist several successful attempts to solve them with computer programs. We solve text-based CAPTCHAs with a deep convolutional neural network that is based on VGG-16. Our network initially is trained with labeled samples and then proceeds with new and harder unlabeled ones. These harder CAPTCHAs have a line through the text or use multiple or different fonts. We show that a network that was trained on one type of CAPTCHAs can adapt to new and more challenging CAPTCHAs even if the network is only using already correctly solved CAPTCHAs. We also show that active learning can improve the accuracy of our networks and decrease the number of learning iterations necessary for them to adapt to more challenging CAPTCHAs.