Handwriting Recognition

1. Dataset

1.1. List of Datasets

1.1.1. Japanese

• Hiragana (ETL 8): the main alphabet. Each Hiragana has a corresponding Katakana
• Katakana (ETL 1): Katakana are used for foreign words.
• Kanji (ETL 8): Kanji are Chinese characters that were adopted in Japanese
• Source for ETL: how to unpack using Medium article
• Kuzushiji: Old Japanese Hiragana, but much more complex
  • Source: Kuzushiji-MNIST

1.2. Standarization

• Some general rules for input image standarization
  • Gray Scale (Black and White): only 1 channel, no RGB (3 channel)
  • White background (255) & Character in Black (~0-180) (refer to the normalization)
  • Normalization: images should be in the range
    • 0 (black) to 1 (white): image/255.
    • -1 (black) to 1 (white): either (image/ 127.5 -1) or -(image/ 127.5 -1) to convert to white background & black character (Depend on the dataset)

10 classes of Kuzushiji-MNIST

2 Conv-RNN-CTC

2.1. CTC

• Explanation of Connectionist Temporal Classification

Resources

Conv-RNN-CTC network

• Build a Handwritten Text Recognition System using TensorFlow
• Handwritten Text Recognition with TensorFlow
• Pytorch Repo

Japanese Multi Character Recognition

• State-of-the-art image classification networks for hand-written hiragana characters

Region Proposal Network (RPN)

• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (6 Jan 2016). Gives a state of the art at the time about RPNs. And explains how anchor based RPNs work

Potentially useful links and tools for RNNs

URLs

• Blog page that suggests a model structure that could prove useful to us.

• Courses for a deep learning bootcamp that ended up on a "Build and deploy and end-to-end deep learning system".

• Tensorflow tutorial that implements and trains a text generator. Really clear for understanding what is going on with RNNs.

• Image annotation tool implemented by Oxford.

Full Stack Deep Learning

• Full Stack Deep Learning

Conv-RNN-CTC network

• Build a Handwritten Text Recognition System using TensorFlow
• Handwritten Text Recognition with TensorFlow
• Pytorch Repo
• Code gathering for everything we try to do !

Japanese Multi Character Recognition

• State-of-the-art image classification networks for hand-written hiragana characters

Region Proposal Network (RPN)

• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (6 Jan 2016). Gives a state of the art at the time about RPNs. And explains how anchor based RPNs work

• Region Proposal Network -- A detailed view. Article that better explains how RPNs work and what are anchors.