Optimized Recognition Of CAPTCHA Through Attention Models

Information retrieval from the CAPTCHA is a crucial part, this CAPTCHA always contains some unwanted information along with required information, so attention technique comes in handy to select such useful information discarding the unwanted part. The attention concept has become a very important part in the field of deep learning which uses Natural Language Processing(NLP) and Computer Vision(CV). Attention mechanism is rigorously used in OCR based applications which requires generating of selected information rather than every information available. Our work includes implementation of general, global and local Attention mechanisms used with two different models which were transfer learning model and the parameter search model. As OCR with attention technique is computationally costly it is required to optimize the entire process so we suggest optimized retrieval of information from CAPTCHA using parameter search algorithm. This retrieval includes using weights that reduced the training time from 4.03 minutes to 3.33 minutes and the number of training images which were used for training were reduced than before. We obtained the highest accuracy of 87.34% for general attention with parameter search model and local attention model with parameter search model proved to have less computation and less training time than the general attention with parameter search model.