EADST

I read some data augmentation papers this week.

Data augmentation has three main areas.

• Space transform
• Color change
• Information delect

Here are four papers where three papers using space transform and one paper taking information delect.

1. Bhunia, Ayan & Das, Abhirup & Bhunia, Ankan & Perla, Sai & Roy, Partha. (2019). Handwriting Recognition in Low-Resource Scripts Using Adversarial Learning. 10.1109/CVPR.2019.00490.

   • They propose the algorithm, Adversarial Feature Deformation Module (AFDM) inspired by Spatial Transformation Networks (STN).

     • Localisation Network: using Generative Adversarial Networks (GANs) to generate the transform matrix.
     • Grid Generator: transforming feature maps with matrix.
     • Sampler: based on the neighbor relative position to update weights.

     

2. Luo, Canjie & Zhu, Yuanzhi & Jin, Lianwen & Wang, Yongpan. (2020). Learn to Augment: Joint Data Augmentation and Network Optimization for Text Recognition.

   • They combine moving least squares with a learnable agent to augment data.

     Code

     

3. C. Wigington, S. Stewart, B. Davis, B. Barrett, B. Price and S. Cohen, "Data Augmentation for Recognition of Handwritten Words and Lines Using a CNN-LSTM Network," 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), Kyoto, 2017, pp. 639-645, doi: 10.1109/ICDAR.2017.110.

   • They reshape the characters with the normal distribution where the parameters from the normalization step.

     

4. Pengguang Chen. GridMask data augmentation. arXiv preprint arXiv:2001.04086, 2020. 3.

   • They use grid mask strategy to shadow some blocks with the grid for image.

   • A handwritten recognition is used in this paper: GridMask Based Data Augmentation for Bengali Handwritten Grapheme Classification

     Code