Research map / Architectures and efficiency / RNNs and normalization
Gated Recurrent Convolution Neural Network for OCR
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 gate to RCL (GRCL instead of RCL) (ablation) | GRCNN vs RCNN · IIIT5K and SVT | accuracy: e.g. 80.21 vs 78.94 on IIIT5K at 3 iters; 80.98 vs 79.76 on SVT | Helped |
| increase GRCL iterations from 1 to 3 (ablation) | GRCNN · IIIT5K and SVT | accuracy: IIIT5K 77.92 -> 80.21; SVT 78.67 -> 80.98 | Helped |
| replace GRCL with plain two-conv layers (no recurrence) (ablation) | feed-forward CNN · IIIT5K and SVT | accuracy: plain 77.21 vs GRCNN(1 iter) 77.92 on IIIT5K | Hurt |
| add peephole connections to BLSTM (all gates, LSTM-A) (ablation) | GRCNN-BLSTM sequence modeling · IIIT5K and SVT | accuracy: LSTM-A 76.52 vs no-peephole 77.92 on IIIT5K | Hurt |
| add peephole to input gate only (LSTM-I) (ablation) | BLSTM · IIIT5K and SVT | accuracy: 76.84 vs 77.92 on IIIT5K | Hurt |
| add peephole to forget gate only (LSTM-F) (ablation) | BLSTM · IIIT5K and SVT | accuracy: 77.26 vs 77.92 on IIIT5K | Hurt |
| add peephole to output gate only (LSTM-O) (ablation) | BLSTM · IIIT5K and SVT | accuracy: 76.91 vs 77.92 on IIIT5K; 78.65 vs 78.67 on SVT | Hurt |
| replace ResNet-20 feature extractor with GRCNN (T=5) in BLSTM framework | GRCNN-BLSTM vs ResNet-BLSTM · IIIT5K, SVT, IC03 | accuracy: IIIT5K 80.8 vs 79.2; SVT 81.5 vs 80.2; IC03 91.2 vs 89.9 | Helped |
| GRCNN-BLSTM (T=5) trained on Synth90k, no fine-tuning, compared to prior methods | GRCNN-BLSTM · IIIT5K, SVT, IC03 lexicon-free/constrained | accuracy: beats CRNN on IIIT5K-1k (95.6 vs 94.4) and IC03-Full (97.8 vs 97.6) but trails on SVT-50 (96.3 vs 96.4) and IC03-50 (98. | Mixed |
Closest papers by meaning
- Recursive Recurrent Nets With Attention Modeling for OCR in the Wild (CVPR 2016 · similarity 0.91)
- Reading Scene Text in Deep Convolutional Sequences (AAAI 2016 · similarity 0.85)
- SqueezedText: A Real-Time Scene Text Recognition by Binary Convolutional Encoder-Decoder Network (AAAI 2018 · similarity 0.85)
- Char-Net: A Character-Aware Neural Network for Distorted Scene Text Recognition (AAAI 2018 · similarity 0.84)
- Learning long-range spatial dependencies with horizontal gated recurrent units (NeurIPS 2018 · similarity 0.84)
- UPOCR: Towards Unified Pixel-Level OCR Interface (ICML 2024 · similarity 0.83)
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.