Domain Adaptive Scene Text Localization in Night View Images

Abstract

Scene text localisation in a night is a challenging problem and a less explored problem in the area of robust reading. The day and night images have a different distribution. Hence, a domain shift exists from day to night images. This paper introduces a domain adaptation network (DAN) to learn the domain shift from day images to night images. The proposed sub-network DAN can be attached to any existing state-of-theart text detector to learn the domain shift from day to night. For the training of the DAN, synthetic data generation has been done by utilising the generative adversarial network. The proposed method has been extensively validated on the public as well as on the synthetic datasets. The proposed DAN improves the performance of the underline text localisation model with a margin of 1.4-4.0%, 0.1-1.0%, and 0.1-3.1% on LP Night Dataset, ICDAR2015, and Total-Text benchmarking datasets.