# 討論 Week18 的主題
###### tags: `OCR track`
1. 反光偵測模型實作
2. text detection 新演算法實作
3. 訓練新的 OCR model
4. hand-writing OCR model
5. 分類模型
6. 其他想法

## RPA support CV 任務

* 昊中：CRNN + CTC (Pytorch)(用cc_ws_ocr專案的dataset去訓練)
* 信賢：data augmentation, OCR, yolo backbone (3 選 1)
* 沛筠：data augmentation
* RPA 任務：驗證碼 OCR
## OCR 步驟

* 分類模型
* efficientNet
* mobilenet
* 影像前處理
* 去噪
* data augmentation
* detection
* text (e.g., craft)
* object (e.g., yolov4)
* segmentation
* OCR
* 前處理：e.g. 文字加深加黑再套用 OCR model
* 精進舊版本的 recogntion model
* hand-writing OCR model
* 信用卡自扣申請書
* 排除
* crnn+ctc: 90 ~ 91%
* [clova](https://github.com/clovaai/deep-text-recognition-benchmark)
* attention
* 支票手寫
* 匯款單
## 執行方式

* :star: 報告順序:
* week 18: 固定演算法，嘗試多種 data augmentation 對 OCR performance 的貢獻 (固定 CRNN+CTC, 參考 if_cv_dev)
* week 21: 先找幾個OCR演算法(except CRNN+CTC、attention、clova)，使用相同的資料去比較 performance
* 後續比較方式:
* case 1: CRNN + CTC & 原始資料 (baseline)
* case 2: hand-writing OCR model & 原始資料
* case 3: CRNN + CTC & data augmentation (method 1, 2, ...)
* case 4: hand-writing OCR model & data augmentation

### augmentation 筆記

## 方法論

信賢：
* :star:Image-to-Image Translation with Conditional Adversarial Networks
- [blog](https://kaonashi-tyc.github.io/2017/04/06/zi2zi.html)
- [paper](https://arxiv.org/pdf/1611.07004.pdf)
- [code1](https://github.com/kaonashi-tyc/zi2zi)
- [code2](https://github.com/jasonlo0509/Font2Font)
* TextRecognitionDataGenerator
[code](https://github.com/Belval/TextRecognitionDataGenerator)
* Text Renderer
[code](https://github.com/Sanster/text_renderer)
* Augmentor
[code](https://github.com/mdbloice/Augmentor)

昱睿：
* :star:AutoAugment: Learning Augmentation Strategies from Data
* paper: [AutoAugment: Learning Augmentation Strategies from Data](https://arxiv.org/pdf/1805.09501.pdf)
* code: [DeepVoltaire/AutoAugment](https://github.com/DeepVoltaire/AutoAugment)

立晟：
* :star:imgaug
- [docs](https://imgaug.readthedocs.io/en/latest/source/examples_basics.html)
* :star:Albumentations
- [paper](https://arxiv.org/pdf/1809.06839.pdf)
- [docs](https://albumentations.readthedocs.io/en/latest/)

昊中：
* imgaug
* https://github.com/Belval/TextRecognitionDataGenerator
* https://github.com/Sanster/text_renderer
* https://github.com/williamyang1991/TET-GAN
* https://github.com/pyliaorachel/cross-lingual-font-style-transfer
* (備)https://github.com/Canjie-Luo/Text-Image-Augmentation
* :star:https://github.com/RubanSeven/Text-Image-Augmentation-python

沛筠：
* :star:RandAugment: Practical automated data augmentation with a reduced search space
* paper: [RandAugment: Practical automated data augmentation with a reduced search space](https://arxiv.org/pdf/1909.13719.pdf)
* code: [pytorch-randaugment](https://github.com/ildoonet/pytorch-randaugment)
* code: [imgaug (現成)](https://github.com/aleju/imgaug/blob/master/imgaug/augmenters/collections.py)
* 為 pytorch 版本的延伸
* code: [tensorflow (較舊)](https://github.com/tensorflow/tpu/blob/1ebfc4b5fa6b5fad0bd7d422832ae8826e8059f2/models/official/efficientnet/autoaugment.py)