Research map / Label-efficient and robust learning / Domain adaptation and generalization
Sequence-To-Sequence Domain Adaptation Network for Robust Text Image Recognition
Yaping Zhang, Shuai Nie, Wenju Liu, Xing Xu, Dongxiang Zhang, Heng Tao Shen
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add gated attention similarity (GAS) to the sequence recognizer (ablation) | SSDAN · IAM test | WER: 23.9 -> 22.2 | Improved |
| Add gated attention similarity (GAS) to the sequence recognizer (ablation) | SSDAN · CROHME 2014 | ExpRate: 39.9 -> 41.6 | Improved |
| Add gated attention similarity (GAS) to the sequence recognizer (ablation) | SSDAN · IIIT5K, SVT, IC-03, and IC-13 | accuracy: 81.1 -> 83.8; 82.1 -> 84.5; 91.2 -> 92.1; 91.0 -> 91.8 | Improved |
| Align CNN outputs directly rather than attended character-level features (ablation) | SSDAN on IAM · IAM test | WER: 23.9 -> 25.0 | Hurt |
| Align CNN outputs directly rather than attended character-level features (ablation) | SSDAN on IAM · IAM test | CER: 9.2 -> 11.1 | Hurt |
| Use DenseNet instead of VGG or ResNet as the encoder (ablation) | SSDAN on IAM · IAM test | WER: 26.9 -> 22.2; 27.9 -> 22.2 | Improved |
| Use DenseNet instead of VGG or ResNet as the encoder (ablation) | SSDAN on IAM · IAM test | CER: 12.6 -> 8.5; 13.1 -> 8.5 | Improved |
| Use CORAL instead of MMD or adversarial loss for domain-shift measurement (ablation) | SSDAN on IAM · IAM test | WER: 22.7 -> 22.2; 24.6 -> 22.2 | Improved |
| Use CORAL instead of MMD or adversarial loss for domain-shift measurement (ablation) | SSDAN on IAM · IAM test | CER: 8.8 -> 8.5; 10.8 -> 8.5 | Improved |
| Vary the confidence threshold in the attention gate (ablation) | SSDAN on MJSYNTH → SVT · SVT | accuracy | Mixed |
| Vary the domain-adaptation loss weight λ (ablation) | SSDAN on MJSYNTH → SVT · SVT | accuracy | Mixed |
| Add unlabeled target examples alongside labeled training examples (ablation) | SSDAN · IC-03 and IIIT5K | accuracy | Improved |
Closest papers by meaning
- Sequence-to-Sequence Contrastive Learning for Text Recognition (CVPR 2021 · similarity 0.86)
- Robust Scene Text Recognition With Automatic Rectification (CVPR 2016 · similarity 0.85)
- An End-to-End Trainable Neural Network for Image-Based Sequence Recognition and Its Application to Scene Text Recognition (TPAMI 2016 · similarity 0.85)
- Char-Net: A Character-Aware Neural Network for Distorted Scene Text Recognition (AAAI 2018 · similarity 0.85)
- Target Semantics Clustering via Text Representations for Robust Universal Domain Adaptation (AAAI 2025 · similarity 0.85)
- CDistNet: Perceiving Multi-domain Character Distance for Robust Text Recognition (IJCV 2023 · 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.
