Abstract
Detecting arbitrary-shaped text in low-light environments is a significant challenge due to noise, blur, and low contrast. Traditional approaches utilize a two-stage pipeline—Low-Light Image Enhancement (LLE) followed by a standard detector—which often introduces artifacts optimized for human aesthetics rather than machine perception. We present a novel single-stage text detection pipeline that bypasses enhancement steps. By incorporating a training-only Spatial-Constrained Learning Module (SCM), a Dynamic Snake FPN (DSF), and a Text Shaping strategy with Rotated Rectangular Accumulation (TSR), our method achieves state-of-the-art results. We also introduce the LATeD dataset, a comprehensive benchmark for low-light arbitrary-shape text detection. Our approach achieves a 67.1% F1-score on LATeD and remains highly competitive on standard benchmarks like CTW1500.
Introduction
Text detection in the wild is a cornerstone of many computer vision applications, yet performance degrades sharply in low-light conditions. The standard industry response has been to apply image enhancement algorithms as a pre-processing step. However, this "straightforward" solution has two primary flaws. First, LLE algorithms are typically designed to make images visually pleasing to humans, often smoothing over the fine edges and textures that machine learning models rely on for text localization. Second, two-stage pipelines are computationally expensive and prone to error propagation between the enhancement and detection phases.
This paper proposes a single-stage alternative that "sees" text directly in the dark. By focusing on architectural innovations that preserve spatial information and adapt to the unique geometry of text, we demonstrate that pre-processing is not only unnecessary but often counterproductive.
Methodology
Spatial-Constrained Learning Module (SCM)
To address the loss of detail in low-light feature maps, we introduce the SCM. This auxiliary module is utilized exclusively during the training phase. It forces the network to learn the reconstruction of spatial and semantic information that is typically lost during feature map resizing. Because the SCM is discarded during inference, it provides a "training-only" boost to accuracy without increasing the computational load of the deployed model.
Dynamic Snake FPN (DSF)
Text instances, particularly in urban environments, often exhibit slender, tubular, or curved shapes. Standard square convolutional kernels are poorly suited for these geometries. We incorporate Dynamic Snake Convolutions (DSC) into a Feature Pyramid Network (FPN). This allows the model to adaptively adjust its receptive field to follow the winding paths of arbitrary-shaped text, significantly improving detection in complex scenes.
Text Shaping with Rotated Rectangular Accumulation (TSR)
Traditional Non-Maximum Suppression (NMS) is often a bottleneck in text detection, especially for curved or overlapping text. We replace NMS with a strategy based on Farthest Point Sampling (FPS). By modeling text as a series of accumulated rotated rectangles, the TSR module better captures the streamline nature of text lines. This shift not only improves topological accuracy but also increases inference speed from approximately 2 FPS to 10 FPS in our ablation studies.
The LATeD Benchmark
To evaluate performance in realistic settings, we introduce the Low-light Arbitrary-shape Text Detection (LATeD) dataset.
- Scale: 1,500 manually curated images containing 13,923 text instances.
- Diversity: Captured across Sydney and Melbourne, the dataset includes multilingual text (English, Chinese, Japanese, Korean) in diverse scenarios such as billboards, road signs, and posters.
- Challenge: The dataset focuses specifically on the intersection of low-light conditions and arbitrary text geometry.
Experimental Results
Performance in the Dark
On the LATeD dataset, our proposed method achieved an F1-score of 67.1%. This represents a significant margin over the previous state-of-the-art model, BPN++ (59.5%), and outperforms various combinations of LLE enhancers and standard detectors. This confirms that a single-stage model trained on raw dark images can superiorly navigate the "Human vs. Machine" vision gap.
Robustness and Efficiency
Despite its specialization for low light, the model is highly robust in normal lighting conditions, achieving an 86.2% F1-score on the CTW1500 benchmark. Furthermore, the TSR module provided a nearly 5x increase in inference speed compared to our baseline, making the model more viable for practical, real-world deployment.
Conclusion
We have demonstrated that a single-stage approach to low-light text detection is both more efficient and more accurate than traditional two-stage enhancement pipelines. By utilizing spatial constraints during training and adaptive convolutions during inference, our model effectively overcomes the noise and contrast issues inherent in dark environments. Future work will focus on integrating this detection pipeline with recognition engines to create a complete end-to-end "Text Spotter" for low-light conditions and further optimizing inference speeds for real-time video processing.
Summary updated