Research map / Detection and segmentation / CNN features
LRANet: Towards Accurate and Efficient Scene Text Detection with Low-Rank Approximation Network
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LRANet, a regression-based arbitrary-shaped scene text detector that represents text contours as linear combinations of SVD-derived eigenanchors and uses a dual sparse/dense assignment scheme for faster NMS-light inference.
Problem
Parameterized text shape regression for arbitrary-shaped scene text detection lacks text-specific shape information and compact representation, and dense-assignment pipelines suffer from slow NMS, while one-to-one approaches lack supervision and need multiple decoders.
Approach
Low-rank approximation via SVD of a labeled text-contour matrix yields data-driven "eigenanchors"; text contours are reconstructed from a few learned coefficients. A dual assignment scheme trains with dense positives and infers with sparse positives (K=3) to cut NMS cost.
Key results
- LRANet with external pretraining reaches 89.0 F on Total-Text and 87.4 F on CTW1500 at 37.2 FPS on CTW1500 (RTX3090 setting).
- Dual assignment achieves 22.1 FPS vs 14.6 for dense assignment on Total-Text while improving F from 87.6 to 88.1.
- LRA with 14 eigenanchors achieves 98.0 IoU on CTW1500 with fewer parameters than TPS (22) and Fourier (22) representations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine dense and sparse assignment (ablation) | LRANet · Total-Text | F-measure: 87.6 -> 88.1 | Helped |
| combine dense and sparse assignment (ablation) | LRANet · Total-Text | FPS: 14.6 -> 22.1 | Helped |
| add dense assignment supervision to sparse assignment (ablation) | LRANet · Total-Text | F-measure: 86.9 -> 88.1 | Helped |
| increase sparse samples per instance from 1 to 3 (ablation) | LRANet · Total-Text | F-measure: 87.3% -> 88.1% | Helped |
| increase sparse samples per instance from 3 to 5 (ablation) | LRANet · Total-Text | F-measure: 88.1 -> 87.8 | Hurt |
| remove NMS (ablation) | LRANet with K=1 · Total-Text | F-measure: 87.3 -> 86.6 | Hurt |
| increase eigenanchor dimension from 10 to 14 (ablation) | LRA · CTW1500 text contour reconstruction | IoU: 95.1 -> 98.0 | Helped |
| increase eigenanchor dimension from 14 to 18 (ablation) | LRA · CTW1500 text contour reconstruction | F-measure: 86.9 -> 86.9 | No clear effect |
| use contour-driven rather than mask-driven data representation (ablation) | LRA · CTW1500 text contour reconstruction | IoU: 90.8 -> 98.0 | Helped |
| use contour-driven rather than side-driven data representation (ablation) | LRA · CTW1500 text contour reconstruction | IoU: 97.9 -> 98.0 | Helped |
| increase input short side from 608 to 1000 (ablation) | LRANet · Total-Text | F-measure: 85.6 -> 89.0 | Helped |
| increase input short side from 608 to 1000 (ablation) | LRANet · Total-Text | FPS: 48.2 -> 22.1 | Hurt |
Surprising findings
- Using only five LRA coefficients (top-1 eigenanchor excluded, u2 to u6) still outperforms TextRay with 44 parameters (81.1 vs 80.6 F) on CTW1500 per the appendix.
- Removing NMS with K=1 drops F-measure only modestly relative to the sparse-with-NMS setting while speeding up inference to 23.5 FPS, though recall falls to 83.8.
Limitations
- Eigenanchor space is learned from the training contour distribution, so generalization depends on dataset shape statistics (only partly checked via SynthText-150K).
- Sparse assignment without NMS reduces recall (83.8 vs 85.7 on Total-Text), showing a speed-accuracy trade-off.
- Requires resampling contour vertices to a fixed N via cubic spline interpolation.
Experimental setup
Datasets: CTW1500 Total-Text MSRA-TD500 SynthText-150K
Benchmarks: CTW1500 Total-Text MSRA-TD500
Metrics: precision recall F-measure FPS IoU
Baselines: DB DB++ TextBPN TextPMs FSG LeafText TextRay FCENet PCR ABCNet V2 TextDCT TPSNet
Builds on: FCENet TextBPN TPSNet focal loss DB
Closest papers by meaning
- Progressive Contour Regression for Arbitrary-Shape Scene Text Detection (CVPR 2021 · similarity 0.89)
- ContourNet: Taking a Further Step Toward Accurate Arbitrary-Shaped Scene Text Detection (CVPR 2020 · similarity 0.89)
- Look More Than Once: An Accurate Detector for Text of Arbitrary Shapes (CVPR 2019 · similarity 0.89)
- OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning (NeurIPS 2025 · similarity 0.86)
- Kernel Adaptive Convolution for Scene Text Detection via Distance Map Prediction (CVPR 2024 · similarity 0.85)
- Arbitrary Reading Order Scene Text Spotter with Local Semantics Guidance (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.