Deep Learning-Based Screen Text Detection and Recognition for Onboard Maintenance Systems

Due to the numerous alarm entries and dense text on the interface of the onboard maintenance system (OMS), it is still challenging for current algorithm to achieve complete and accurate text recognition based on screen. This paper proposes a deep learning method that employs YOLOv3, CRNN and a post-processing module. This method first uses YOLOv3 to locate some text which we really concerned; then the feature map of the positioning area is aligned and fed to the CRNN text recognition module to obtain the text detection result, and finally the text matching module based on the minimum edit distance is used to solve some hard cases. During the experiment, we also discussed the limits of each module and made some improvements. The experimental results demonstrate that the approach we proposed has an accuracy of 99.95% and a recall of 95.74%, indicating that our algorithm can solve the problems of incomplete localization and inaccurate text recognition, and can meet practical application scenarios.