Abstract: CAPTCHA is a kind of network security mechanism that blocks machines from abusing network resource owned by human. Studying the recognition of CAPTCHA can help to find its hidden defects, and thus make it securer. To read closely-connected CAPTCHA that can hardly be recognized by methods of state of art, this paper brought up a new recognition algorithm based on rejection. During the process of this algorithm, recurrent neural network (RNN) was first used to recognize the unknown CAPTCHA. Then, to make the recognition results reliable, a new rejection algorithm was brought up. Data dimension reduction was also performed on rejection features. Experiment results show the following three points: Firstly, our new recognition algorithm can recognize closely-connected CAPTCHA with high reliability. Secondly, the new rejection algorithm is superior to other methods of state of art. Lastly, data dimension reduction algorithm can improve the performance of the rejection algorithm, thus making the recognition results more reliable.