Research map / Label-efficient and robust learning / Recognition and long-tail
Self-Distillation Regularized Connectionist Temporal Classification Loss for Text Recognition: A Simple Yet Effective Approach
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 DCTC, a module-free self-distillation regularizer for CTC text recognition that estimates frame-wise latent alignments via a closed-form MAP rule computed from CTC gradients and probabilities.
Problem
CTC-based text recognition models lose accuracy because CTC optimizes whole-sequence likelihood and neglects individual character supervision; distillation for CTC suffers from alignment inconsistency and typically needs extra teacher modules.
Approach
Adds a frame-wise cross-entropy term to CTC loss whose target alignment is obtained in closed form as argmin over classes of G/P (CTC gradient divided by output probability), derived from a MAP estimate of the latent alignment, so the model acts as its own teacher with no extra modules, parameters, or data.
Key results
- DCTC raises CRNN average English accuracy from 77.3 to 79.9 (+2.6) and Chinese average from 68.7 to 70.8 (+2.1).
- SVTR-T improves from 90.8 to 91.7 English average and 75.3 to 75.9 Chinese average with DCTC, with no inference cost change.
- Module-dependent alignment estimates (Self, Teacher) degrade accuracy for most models, e.g. CRNN Self drops Chinese average by 7.1 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace CTC loss with DCTC loss | CRNN · English benchmarks (Avg) | accuracy: 77.3 -> 79.9(+2.6) | Helped |
| Replace CTC loss with DCTC loss | CRNN · Chinese benchmarks (Avg) | accuracy: 68.7 -> 70.8(+2.1) | Helped |
| Replace CTC loss with DCTC loss | SVTR-T · English benchmarks (Avg) | accuracy: 90.8 -> 91.7(+0.9) | Helped |
| Replace CTC loss with DCTC loss | SVTR-T · Chinese benchmarks (Avg) | accuracy: 75.3 -> 75.9(+0.6) | Helped |
| Replace CTC loss with DCTC loss | SVTR-S · English benchmarks (Avg) | accuracy: 91.6 -> 92.5(+0.9) | Helped |
| Replace CTC loss with DCTC loss | SVTR-S · Chinese benchmarks (Avg) | accuracy: 76.3 -> 77.3(+1.0) | Helped |
| Replace CTC loss with DCTC loss | SVTR-B · English benchmarks (Avg) | accuracy: 92.3 -> 93.1(+0.8) | Helped |
| Replace CTC loss with DCTC loss | SVTR-B · Chinese benchmarks (Avg) | accuracy: 77.5 -> 78.2(+0.7) | Helped |
| Replace CTC loss with DCTC loss | SVTR-L · English benchmarks (Avg) | accuracy: 92.8 -> 93.3(+0.5) | Helped |
| Replace CTC loss with DCTC loss | SVTR-L · Chinese benchmarks (Avg) | accuracy: 78.1 -> 79.2(+1.1) | Helped |
| Use FocalCTC instead of CTC loss | CRNN · English benchmarks (Avg) | accuracy: 77.3 -> 75.6(-1.7) | Hurt |
| Use EnCTC instead of CTC loss | CRNN · Chinese benchmarks (Avg) | accuracy: 68.7 -> 64.2(-4.5) | Hurt |
Surprising findings
- Estimates from the model's own argmax (Self) or an added Transformer teacher branch often hurt accuracy relative to the CTC baseline.
- EnCTC reduces Chinese accuracy for CRNN by 4.5 points, showing some CTC variants can hurt performance.
Limitations
- Relies on a proof of boundedness supplied only in supplementary material; method derivation has heuristic steps (proportionality assumptions).
- Gains are modest for stronger models (e.g. SVTR-L average +0.5 on English).
- Evaluated only on PaddleOCR-based architectures with fixed hyperparameters from base models; λ is tuned per language setting.
Experimental setup
Datasets: ST MJ Chinese Benchmark (Scene, Web, Document, Handwritten)
Benchmarks: IC13 IC15 SVT SVTP IIIT CT Chinese Benchmark (Scene, Web, Doc, Hand)
Metrics: accuracy alignment accuracy (AACC)
Baselines: CTC loss FocalCTC EnCTC Self Teacher
Builds on: CTC CRNN SVTR FocalCTC EnCTC
Closest papers by meaning
- CR-CTC: Consistency regularization on CTC for improved speech recognition (ICLR 2025 · similarity 0.87)
- Align With Purpose: Optimize Desired Properties in CTC Models with a General Plug-and-Play Framework (ICLR 2024 · similarity 0.85)
- Accurate Scene Text Recognition with Efficient Model Scaling and Cloze Self-Distillation (CVPR 2025 · similarity 0.84)
- Rethinking Self-Distillation: Label Averaging and Enhanced Soft Label Refinement with Partial Labels (ICLR 2025 · similarity 0.83)
- Retaining Knowledge and Enhancing Long-Text Representations in CLIP through Dual-Teacher Distillation (CVPR 2025 · similarity 0.83)
- On the Generalization of Handwritten Text Recognition Models (CVPR 2025 · similarity 0.82)
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.