Abstract
Traffic sign text detection is a vital component of autonomous driving systems, providing contextual information beyond simple sign recognition. However, environmental factors like low lighting and the geometric complexity of long, sparse text strings pose significant challenges. This paper presents HTSTL (Head-and-Tail Search Network), a robust detection pipeline featuring a Scale-Transfer Layer (STL) for efficient resolution recovery and a novel Head-and-Tail pixel classification strategy. Experimental results on the CTST-1600 dataset demonstrate that HTSTL achieves state-of-the-art performance (F-score of 0.82) while maintaining real-time processing speeds.
Introduction
Autonomous vehicles must interpret complex road environments in real-time. While object detection for standard traffic signs is well-established, "reading" the text on these signs remains difficult. Challenges include poor visibility at night, the extreme aspect ratios of Chinese traffic text, and the limited computational resources available on-board vehicles. Existing word-based detectors often fragment long text lines or fail in low-contrast scenarios.
Methodology
The HTSTL framework addresses these issues through a three-stage process: preprocessing, feature extraction via a modified Fully Convolutional Network (FCN), and structural post-processing.
Preprocessing and Scale-Transfer
To mitigate low-light failures, a preprocessing module applies Gamma correction based on the image's Root Mean Square (RMS) brightness. The core of the FCN architecture is the Scale-Transfer Layer (STL). Unlike traditional upsampling methods that introduce zero-padding or learnable parameters, STL rearranges channel-wise data into spatial dimensions. This approach restores resolution efficiently, reducing computational overhead.
Head-and-Tail Search Strategy
To handle long and sparse text, the network classifies pixels into three distinct categories: Head, Tail, and Middle. This structural classification allows the system to identify the boundaries of a text instance and link the internal components reliably, preventing the fragmentation common in standard bounding-box regression. By finding the "ends" of a sign and filling in the middle, the model treats detection as a structural search rather than a simple geometric exercise.
Experimental Results
The model was evaluated on the CTST-1600 dataset, specifically curated for Chinese traffic signs, as well as standard benchmarks like ICDAR 2013 and MSRA-TD500.
- Accuracy: HTSTL achieved an F-score of 0.82 on CTST-1600, significantly higher than the EAST (0.76) and TextSnake (0.64) models. On the ICDAR 2013 benchmark, it reached a precision of 0.92.
- Efficiency: The system operates at approximately 5.28 FPS. While slightly slower than the EAST detector (6.33 FPS), it offers superior robustness in the complex scenarios for which it was designed.
- Robustness: Qualitative results confirm the model's ability to detect text in near-total darkness where baseline models often fail.
Conclusion and Future Work
HTSTL provides a specialized solution for traffic sign text detection by combining efficient architectural components with a structurally-aware classification strategy. While highly effective for linear text in diverse lighting, future improvements will target curved text and overlapping instances. The authors also plan to integrate a text recognizer (OCR) directly into the pipeline to create an end-to-end reading system for autonomous navigation.
Summary updated