Rail Transit Line-Sign Text Detection With Patch-Based Region Proposal Network

Rail transit line-sign text detection (RTLSTD) plays a crucial role in rail transit maintenance, yet detecting marking text on rail transit line-signs (RTLS) remains challenging due to the large number of small-scale samples, complex background noise, motion blur, and low-contrast in rail transit scenarios. Despite these challenges, RTLSTD has received little attention due to data collection difficulties. In this paper, we propose a novel rail transit line-sign text detection network for robust detection of line-sign text in rail transit scenes. Specifically, we first design a patch-based region proposal network to generate candidate text regions, which alleviates noise interference from the rail transit environment by dividing feature maps into feature patches. Then, a multi-scale feature enhancement module with mask attention is designed to enhance the model’s representation capability for diverse rail transit line-sign samples. Furthermore, we address the motion blur challenge in RTLSTD by employing multi-scale feature learning and frequency modulation. To promote research on RTLSTD, we contribute a real-world rail transit line-sign text dataset with various challenges, along with a synthetic dataset specifically addressing motion blur challenges. Extensive experiments on our datasets demonstrate that our proposed method achieves state-of-the-art performance. The code is available online at https://github.com/ChengliZhu777/RTLST