Introduction
Detecting text in natural scenes—such as street signs, product packaging, or curved advertisements—presents a significant challenge compared to standard document scanning. Text in the "wild" is often arbitrary-shaped, appearing curved, slanted, or irregularly distorted. Current research typically follows two paths: segmentation-based methods, which are flexible but noise-sensitive, and boundary-point-based methods, which often rely on global modeling that neglects crucial local spatial details.
To address these limitations, we present BPDO (Boundary Points Dynamic Optimization), a two-stage framework designed for high-precision detection of arbitrary-shaped text. The core philosophy of BPDO is that the exact edge of a word is best determined by examining the immediate neighborhood of boundary points rather than the text instance as a whole.
The BPDO Framework
The BPDO architecture operates in two distinct stages to move from a coarse text proposal to a refined, accurate boundary.
Stage 1: Text Boundary Points Proposal
The process begins with a ResNet-50 backbone and Feature Pyramid Networks (FPN) to extract multi-scale visual features. These features are processed by a novel Text Aware Module (TAM). The TAM employs both channel and spatial attention mechanisms to fuse features across different scales, allowing the model to effectively "sense" text regions regardless of their orientation or size. This stage generates a coarse set of boundary points along with auxiliary maps for distance, direction, and classification.
Stage 2: Boundary Points Dynamic Optimization (DOM)
The defining innovation of BPDO is the Dynamic Optimization Module (DOM). Inspired by deformable attention, the DOM iteratively refines the coarse boundary points. Instead of a global view, the module samples features from the immediate local neighborhood of each point. By querying these local features, the DOM calculates precise offsets to nudge each point toward the actual text edge. Empirical results suggest that three iterations are sufficient to achieve high-precision alignment.
Experimental Results
BPDO was rigorously evaluated on several benchmark datasets, including Total-Text (curved text), CTW1500 (long curved text), and MSRA-TD500 (multi-directional English and Chinese text).
Performance Highlights
- MSRA-TD500: BPDO achieved a state-of-the-art (SOTA) F-measure of 91.47%, outperforming the previous best model (TextBPN++) by 1.3%.
- CTW1500 & Total-Text: The model reached F-measures of 86.56% and 87.2% respectively, demonstrating robust performance across diverse text geometries.
Ablation studies confirmed the synergy between the TAM and DOM; while the DOM improved the baseline F-measure by 0.46%, the combined architecture yielded a total improvement of 1.37%.
Conclusion and Future Work
BPDO demonstrates that local spatial features are more informative for boundary refinement than global relationships in scene text detection. By focusing on the area immediately surrounding boundary points, the model avoids distraction from distant image features and achieves superior accuracy.
While BPDO sets a new standard for detection, it currently functions as a standalone detector. Future research will focus on evolving this framework into an end-to-end scene text spotting system that integrates both detection and recognition into a single, unified pipeline.
Summary updated