lumimusta/Low-light_Scene_Text_Dataset

Low-light Scene Text Dataset

This repository provides a low-light scene text recognition dataset for studying text recognition under challenging illumination conditions. The dataset is designed to support research on Low-light Scene Text Recognition (LLSTR), where text images may suffer from low contrast, noise, uneven illumination, blur, and other degradations commonly observed in nighttime or poorly lit environments.

The dataset contains two main parts:

• —LSTR: a large-scale low-light scene text recognition dataset derived from well-lit scene text datasets, including ICDAR2015, IIIT5K, and WordArt.
• —ESTR: a real-world nighttime scene text evaluation set containing real low-light street-scene images with English and Spanish text.

The goal of this dataset is to provide a benchmark for evaluating OCR and low-light image enhancement methods in dark environments, especially for approaches that aim to preserve text readability rather than only improve visual brightness.

Dataset Files

This repository contains the following files:

Splits

The dataset is organized into the following splits:

Dataset Description

Low-light scene text recognition is challenging because images captured in dark environments often contain weak contrast, noise, color distortion, and uneven illumination. These degradations can significantly reduce the performance of standard OCR models trained mainly on well-lit images.

This dataset provides low-light and well-lit scene text images for training and evaluation. The low-light portion is based on scene text images from ICDAR2015, IIIT5K, and WordArt, while the real-world evaluation subset contains nighttime street-scene images collected under natural low-light conditions.

The real-world subset includes diverse text appearances, backgrounds, fonts, and languages, with text instances in English and Spanish.

Dataset Statistics

The dataset is intended as a compact benchmark for low-light text recognition research. It is especially useful for evaluating whether OCR systems and low-light enhancement methods can preserve fine text structures such as character strokes, edges, and boundaries.

Usage

After downloading the dataset, extract the image archives before using the annotation files.

Example:

unzip low_light_train.zip -d low_light_train
unzip well_lit_train.zip -d well_lit_train
unzip low_light_test.zip -d low_light_test
unzip well_lit_test.zip -d well_lit_test
unzip low_light_test_real.zip -d low_light_test_real

The annotation files can then be loaded from the corresponding .jsonl files.

Example:

import json

with open("train.jsonl", "r", encoding="utf-8") as f:
    samples = [json.loads(line) for line in f]

print(samples[0])

Intended Use

This dataset is intended for research on:

• —Low-light scene text recognition
• —Robust OCR under nighttime or poorly illuminated conditions
• —Low-light image enhancement for text readability
• —Joint optimization of image enhancement and OCR models
• —Benchmarking OCR models under illumination degradation

Citation

If you use this dataset, please cite the corresponding paper:

@inproceedings{fu2026reading,
  title={Reading in the Dark: Low-light Scene Text Recognition},
  author={Xuanshuo Fu and Lei Kang and Ernest Valveny and Dimosthenis Karatzas and Javier Vazquez-Corral},
  booktitle={ICPR},
  year={2026}
}

Notes

The image files are stored as ZIP archives. The Hugging Face Dataset Viewer may only display the .jsonl annotation files and may not preview all images inside the ZIP archives directly. Please download and extract the ZIP files for full dataset usage.