Research map / Detection and segmentation / CNN features
CPN: Complementary Proposal Network for Unconstrained Text Detection
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 CPN, which runs a segmentation-based semantic proposal branch (Deformable MSN with differentiable deformable morphological dilation) and an anchor-based Balanced RPN in parallel, fused via Interleaved Feature Attention, for scene text detection.
Problem
Scene text detection methods are either segmentation-based (good for irregular shapes but weak on compact/overlapping layouts and needing CPU post-processing) or anchor-based (good for complex layouts but poor on irregular shapes and extreme aspect ratios).
Approach
Combine semantic proposals from a GPU-parallel deformable morphology network with geometric proposals from an anchor-based RPN with a reduced proposal count, and let their features interact via interleaved spatial and scale attention before RRoI-align and the RoI head.
Key results
- Achieves 81.7% f-measure on IC19-ArT, a 3.6% gain over prior state of the art, with a 6.2% recall gain.
- Reaches 90.4% f-measure on IC15 at 11.2 FPS, outperforming FSG by 1.3% and running faster than Mask R-CNN (9.6 FPS).
- Achieves 80.0% f-measure on IC17-MLT and 88.8% on CTW1500; IFA adds 0.5% (IC17-MLT) and 0.9% (CTW1500) F-measure while cutting GFLOPs by about 15%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CPN instead of FSG | Scene text detection · IC15 | F-measure: 89.1 -> 90.4 | Helped |
| Use CPN instead of TextPMs | Scene text detection · MSRA-TD500 | F-measure: 88.9 -> 89.9 | Helped |
| Use CPN instead of DPText-DETR | Scene text detection · CTW1500 | F-measure: 88.8 -> 88.8 | No clear effect |
| Use CPN instead of DPText-DETR | Scene text detection · IC19-ArT | F-measure: 78.1 -> 81.7 | Helped |
| Use CPN instead of FSG | Scene text detection · IC17-MLT | F-measure: 79.6 -> 80.0 | Helped |
| Add Balanced RPN to Deformable MSN (ablation) | CPN · IC17-MLT | F-measure: 72.7 -> 79.5 | Helped |
| Add Balanced RPN to Deformable MSN (ablation) | CPN · CTW1500 | F-measure: 87.1 -> 87.9 | Helped |
| Add Deformable MSN to Balanced RPN (ablation) | CPN · IC17-MLT | F-measure: 78.6 -> 79.5 | Helped |
| Add Deformable MSN to Balanced RPN (ablation) | CPN · CTW1500 | F-measure: 87.3 -> 87.9 | Helped |
| Add IFA (ablation) | CPN with both proposal branches · IC17-MLT | F-measure: 79.5 -> 80.0 | Helped |
| Add IFA (ablation) | CPN with both proposal branches · CTW1500 | F-measure: 87.9 -> 88.8 | Helped |
| Vary the geometric proposal count from 300 to 500, 1000, or 2000 (ablation) | CPN · IC15 | F-measure: 90.36; 90.40; 90.41; 90.41 | No clear effect |
Surprising findings
- Balanced RPN achieves higher recall than Deformable MSN on IC17-MLT, suggesting anchor-based methods may beat segmentation-based ones on challenging multilingual data.
- Using 100 geometric proposals already gives most of the performance; increasing to 2000 adds little, so the semantic branch reduces reliance on many proposals.
Limitations
- Segmentation-based semantic branch inherits sensitivity to complex layouts; authors note anchor branch can be stronger on some datasets.
- Hardware used for training and timing is not specified, limiting comparability of FPS numbers.
Experimental setup
Datasets: IC17-MLT IC19-ArT IC15 MSRA-TD500 CTW1500 SynthText
Benchmarks: IC15 MSRA-TD500 CTW1500 IC19-ArT IC17-MLT
Metrics: precision recall f-measure FPS GFLOPs parameters
Baselines: PSENet-1s CRAFT LOMO PCR DB++ TextBPN TextPMs DRRG Contour I3CL FSG DPText-DETR
Builds on: Mask R-CNN Deformable convolution DBNet Oriented RPN
Closest papers by meaning
- ContourNet: Taking a Further Step Toward Accurate Arbitrary-Shaped Scene Text Detection (CVPR 2020 · similarity 0.87)
- Deep Relational Reasoning Graph Network for Arbitrary Shape Text Detection (CVPR 2020 · similarity 0.87)
- Scene Text Detection with Supervised Pyramid Context Network (AAAI 2019 · similarity 0.86)
- EAST: An Efficient and Accurate Scene Text Detector (CVPR 2017 · similarity 0.86)
- Multi-Scale FCN With Cascaded Instance Aware Segmentation for Arbitrary Oriented Word Spotting in the Wild (CVPR 2017 · similarity 0.85)
- LayoutFormer: Hierarchical Text Detection Towards Scene Text Understanding (CVPR 2024 · similarity 0.84)
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.