A Text Correction and Recognition for Intelligent Railway Drawing Detection

Abstract

In current train control system, there are many drawings that need to be identified manually. This approach leads to many problems, such as misidentification and low efficiency. It is difficult to recognize these drawings automatically because the text on them often has a series of changes, like rotation, tilt, font changes and close to lines. To solve this problem, we divide the task into two parts: text recognition and graphical symbol recognition. This article focuses on text recognition. We use aerial detection technology to classify and detect graphical symbol and text, followed by choosing BILSTM to conducting sequence modeling and using convolutional recurrent neural network (CRNN) iterative training to focus on single word rotation, including tilting, noise-addition, and blurring post-processing. So that the training model can cope with complex scenario and improve the recognition rate of text on drawings. Finally, RESNET is applied to CRNN feature extraction network, and CRNN outputs the recognition and detection results in sequence according to the detection sequence, achieving entry-level detection, and the text recognition rate reaches 98.36%