Robust Text Detection in Foggy Traffic Scenes Using an Enhanced CTPN Model with De-fogging Pre-processing

Abstract Text detection in natural scenes, particularly under challenging weather conditions like fog, remains a formidable task due to complex backgrounds, irregular text arrangements, and uneven illumination. This paper introduces a robust text detection method tailored for foggy traffic images, leveraging an improved Connectionist Text Proposal Network (CTPN) model. To enhance text clarity in foggy conditions, a defogging pre-processing step is incorporated, inspired by atmospheric scattering models. Additionally, a hybrid post-processing module combining NMS and Soft-NMS is proposed to refine text detection results, particularly for horizontal texts. Experimental evaluations on the ICDAR2013 and HSText-1000 datasets demonstrate the superior performance of our method, achieving recall, precision, and F-score improvements of 16.13%, 2.96%, and 10.16%, respectively, over the original CTPN model on the HSText-1000 dataset. This work provides a promising architecture for text detection in adverse weather conditions.