CAPTCHA Recognition Using Deep Convolutional Neural Networks (DCNN)

Abstract

This paper evaluates on captcha recognition using deep convolutional neural network. CAPTCHA is completely automated computer program which is used to publicly test and recognise human and robot separately. The captcha software builds very ambiguous pattern of alpha-numeric to protect the site being accessed by the robots, but these captchas maybe tough to crack by humans too. By using the deep learning model which can identify the captcha by studying the pattern of alphabets and numbers. This method can be further be used for handwriting recognition and number plate recognition. The major key feature in order to decipher the captcha was the pattern of the alphabet and the number, in the deep learning the convolution neural networks are the best to approach this. In convolutional neural network the feature maps help in identifying or recognizing the feature. To render the best outcome deep convolutional neural networks (DCNN) are used. As the deep CNN has many layers efficient and effective computation is done. A simple CNN can do the work but here the captcha has a very confusing forms which may not be effectively solved by the CNN alone. The convolutional operations performed on the sub matric of the picture of a captcha with the feature map matrix which has less index than the sub matric, the sub matric which accounts average convolutional value 1 is said to be the feature which is aimed to be recognized. As deep CNN has comparatively more layers it can compute the model deeply. This method helps the robots not identify the captchas and humans can use the deep learning models to solve it, and the robots are unaware of it. Hence, the DCNN method was the most efficient way to achieve the objective of CAPTCHA recognition with the required accuracy and minimal loss.