CAPTCHA Recognition with Active Deep Learning
F. Stark, Caner Hazırbaş, Rudolph Triebel, Daniel Cremers · 2015
Abstract. CAPTCHAs are automated tests to tell computers and humans apart. They are designed to be easily solvable by humans, but unsolvable by machines. With Convolutional Neural Networks these tests can also be solved automatically. However, the strength of CNNs relies on the training data that the classifier is learnt on and especially on the size of the training set. Hence, it is intractable to solve the problem with CNNs in case of insufficient training data. We propose an Active Deep Learning strategy that makes use of the ability to gain new training data for free without any human intervention which is possible in the special case of CAPTCHAs. We discuss how to choose the new samples to re-train the network and present results on an auto-generated CAPTCHA dataset. Our approach dramatically improves the performance of the network if we initially have only few labeled training data. 1