Position-Encoding Convolutional Network to Solving Connected Text Captcha

Abstract Text-based CAPTCHA is a convenient and effective safety mechanism that has been widely deployed across websites. The efficient end-to-end models of scene text recognition consisting CNN attention-based RNN show limited performance in solving text-based CAPTCHAs. In contrast with the street view image document, character sequence non-semantic. loses its ability to learn semantic context only implicitly encodes relative position extracted features. Meanwhile, security features, which prevent characters from segmentation recognition, extensively increase complexity this model sensitive different schemes. paper, we analyze properties accordingly consider it as highly position-relative task. We propose network named PosConv leverage information without RNN. uses novel padding strategy modified convolution, explicitly encoding into local features characters. This makes CAPTCHAs more informative robust. validate on six schemes, achieves state-of-the-art or competitive accuracy significantly fewer parameters faster convergence speed.