Abstract
TextRay is a single-shot, anchor-free detector for arbitrary-shaped scene text that models text contours globally using a polar coordinate system and Chebyshev polynomial approximation. By predicting a unified geometric encoding instead of relying on complex bottom-up pixel grouping, it simplifies the detection pipeline and achieves competitive accuracy with just a single Non-Maximum Suppression (NMS) post-processing step.
Introduction
Detecting text in natural scenes is notoriously difficult because text instances can appear in wild, arbitrary shapes—they can be highly curved, randomly rotated, or possess extreme aspect ratios. Most state-of-the-art methods tackle this using a "bottom-up" approach, decomposing text into local units (like individual pixels or small bounding box segments) and using complex, heuristic-heavy post-processing to stitch them back together. This process is computationally heavy and sensitive to noise. TextRay addresses this by proposing a "top-down" network that directly learns and predicts the entire complex geometric contour of a text instance in a single pass.
Methodology
TextRay introduces a novel geometric modeling approach embedded in a fully convolutional, single-shot network using ResNet50 and a Feature Pyramid Network (FPN).
Polar System Modeling and Chebyshev Approximation
Instead of predicting variable-length Cartesian coordinates, TextRay constructs a polar coordinate system centered on the text. It emits a fixed number of rays outward to intersect the text contour, capturing the shape as a 1D function of radius versus angle, $r(\theta)$.
To compress these radiuses and capture underlying geometric patterns, the network fits the radius function using Chebyshev polynomials. The resulting fixed-length polynomial coefficients act as a unified "shape vector" (geometric encoding) that the network regresses.
Training Strategies
- Content Loss: Directly regressing abstract polynomial coefficients can be unstable. TextRay uses a bidirectionally differentiable mapping scheme: it converts predicted coefficients back into the radius shape space and calculates the loss against the ground-truth radiuses.
- Central-Weighted Training: For long text instances, points near the ends lack global context. TextRay assigns higher training weights to positive points located closer to the text center, ensuring the network learns from the most context-rich features.
Experiments and Results
The authors evaluated TextRay on three standard arbitrary-shaped scene text benchmarks: SCUT-CTW1500, TotalText, and ICDAR-ArT.
Key Performance Metrics
TextRay achieved highly competitive results as a light-weighted, one-stage detector:
- SCUT-CTW1500: 81.56% F-measure (82.80% Precision, 80.35% Recall).
- TotalText: 80.59% F-measure (83.49% Precision, 77.88% Recall).
Ablation studies confirmed the effectiveness of the proposed components. On SCUT-CTW1500, Content Loss improved the F-measure by 4.12%, and Central-Weighted Training added another 2.91%. Combined, they provided a 5.58% boost.
Conclusion
TextRay demonstrates that mathematical modeling can significantly simplify detection architectures. By converting complex 2D polygons into 1D polar functions, the task becomes a clean, fixed-length regression problem. However, the polar approach has limitations with severely non-convex contours (e.g., tight "U" shapes), where a single ray might intersect the contour multiple times. Future work aims to address these complex global modeling challenges.
Summary updated