Low-Light Scene Text Image Enhancement in the Wild

Abstract

In recent years, the low-light image enhancement task has aroused widespread attention in the image enhancement community since it benefits many downstream tasks, such as scene understanding and nighttime detection. Most existing low-light image enhancement methods focus on general scene images while ignoring the commonly-seen texts in natural scenes. However, to better perform scene understanding in the low-light case, more readable texts in enhanced scene images are necessary. To this end, we propose a real L ow-l I ght S cene T ext image enhancement dataset, named LIST . Similar to existing low-light image enhancement datasets, samples in LIST are pairs of normal-light and low-light scene text images, which are captured with different exposure time. LIST is a bi-lingual low-light scene text image enhancement dataset, containing both Chinese and English. In addition, we design a Text Image Lighting Network, called TILN to transform the low-light scene text images into normal-light ones. TILN is composed of two branches: text image lighting branch and pluggable text aligning branch. The text image lighting branch is used to extract deep semantic features and further enhance the input low-light images; the pluggable text aligning branch leverages a pre-trained character recognition model to obtain glyph representations of texts, which are then utilized to align with the visual features of low-light text images, thereby enhancing its text perception capabilities. Extensive experiments are conducted on LIST, and the results validate the effectiveness of TILN.