Seeing Text in the Dark: Algorithm and Benchmark

TL;DR

This work tackles robust arbitrary-shaped text detection in extremely low-light conditions by moving beyond the conventional enhance-then-detect paradigm. It introduces a one-stage detector augmented with a Spatial-Constrained Learning Module that preserves positional and contextual text information during feature map resizing, and complements it with bottom-up topology-aware components Dynamic Snake FPN and Text Shaping with Rotated Rectangular Accumulation. A dedicated low-light dataset LATeD is presented to benchmark performance across multilingual, curved texts, where the proposed method achieves state-of-the-art results under low-light settings and remains competitive on standard well-lit benchmarks. The combination of constrained training, topology-preserving feature fusion, and efficient contour shaping offers a practical solution for downstream OCR and scene-text understanding in real-world, challenging illumination conditions. The work also provides code and LATeD to facilitate further research.

Abstract

Localizing text in low-light environments is challenging due to visual degradations. Although a straightforward solution involves a two-stage pipeline with low-light image enhancement (LLE) as the initial step followed by detector, LLE is primarily designed for human vision instead of machine and can accumulate errors. In this work, we propose an efficient and effective single-stage approach for localizing text in dark that circumvents the need for LLE. We introduce a constrained learning module as an auxiliary mechanism during the training stage of the text detector. This module is designed to guide the text detector in preserving textual spatial features amidst feature map resizing, thus minimizing the loss of spatial information in texts under low-light visual degradations. Specifically, we incorporate spatial reconstruction and spatial semantic constraints within this module to ensure the text detector acquires essential positional and contextual range knowledge. Our approach enhances the original text detector's ability to identify text's local topological features using a dynamic snake feature pyramid network and adopts a bottom-up contour shaping strategy with a novel rectangular accumulation technique for accurate delineation of streamlined text features. In addition, we present a comprehensive low-light dataset for arbitrary-shaped text, encompassing diverse scenes and languages. Notably, our method achieves state-of-the-art results on this low-light dataset and exhibits comparable performance on standard normal light datasets. The code and dataset will be released.

Figures (20)

• Figure 1: (a) The visual statistics and examples of low-light text images of different scenes, languages, and lighting conditions in the newly curated LATeD dataset. All images are enhanced for clearer vision. (b) Single-stage text detectors originally designed for normal lighting conditions struggle with low-light images. Even with low-light enhancement and fine-tuned with bpn++ following a two-stage step, the results remain unsatisfactory. This is because the enhancer, aimed at overall improvement of visibility, may inadvertently compromise text features.
• Figure 2: The overall structure of the proposed method, where "1/1,256"...indicate the resize ratio and the channel number. The SCM is only employed during the training stage for assisting spatial information awareness of low-light text.
• Figure 4: Text shaping with rotated rectangular accumulation on normal light text images. As shown in the 2nd row, our method remains unaffected by obstructions, showing reliable performance in modeling text streamline features.
• Figure 6: Visualization of text detection results on well-lit datasets (see Supplementary for more visual examples).
• Figure : Input