Using a Simple ACGAN for Manchu Script Dataset Supplementation

Poster presentation: 2023 Korea Multimedia Society Spring Conference

Abstract: The Manchu script is difficult to process with deep learning techniques in part because there is not much data widely available for neural network training. Therefore, this research trained a simple ACGAN for the purpose of supplementing an existing, small-scale Manch dataset. A dataset consisting of a total of 4,000 Manchu script letters was trained in an ACGAN for 30,000 steps. The resulting images produced by the generator model were sufficiently recognizable for dataset supplementation.

Transcript

1. Using a Simple ACGAN
   For Manchu Script
   Dataset Supplementation
   Aaron Snowberger • 이충호, 한밭대학교
   2023.05.19 한국멀티미디어학회 춘계학술대회
   

   View Slide

2. Abstract
   만주어 스크립트는 부분적으로 신경망 훈련에 널리 사용 가능한 데이터가 많지 않기 때문에 딥 러닝
   기술로 처리하기 어렵다. 따라서 본 연구에서는 기존의 소규모 만주어 데이터 세트를 보완하기 위해
   간단한 ACGAN을 훈련하였다. 총 4,000개의 만주 문자로 구성된 데이터 세트는 ACGAN에서 30,000 단계
   동안 훈련되었다. 생성 모델에 의해 생성된 결과 이미지는 데이터 세트 보완을 위해 충분히 인식
   가능하였다.
   The Manchu script is difficult to process with deep learning techniques in part because there is
   not much data widely available for neural network training. Therefore, this research trained a
   simple ACGAN for the purpose of supplementing an existing, small-scale Manch dataset. A
   dataset consisting of a total of 4,000 Manchu script letters was trained in an ACGAN for 30,000
   steps. The resulting images produced by the generator model were sufficiently recognizable for
   dataset supplementation.
   Keywords:
   GAN | ACGAN | Manchu Script | Dataset Supplementation
   

   View Slide

3. Introduction
   Difficulties in Preprocessing Unavailability of Datasets
   Manchu script is written vertically,
   with every letter of a word
   connected by a central stem. This
   makes segmentation of letters
   difficult for pre-processing.
   Large datasets of Manchu script are
   not widely available for machine
   learning. This is the problem this
   research paper addresses.
   

   View Slide

4. Related Research
   Two preprocessing techniques Dataset in this Research
   Two main techniques for Manchu script
   recognition have been used. The first has
   been to attempt letter segmentation as a part
   of the pre-processing step[1,2], but this has
   proven difficult. The second has been to
   attempt segmentation-free recognition[3].
   This research utilizes an existing, small-scale
   dataset of segmented Manchu script letters[4]
   with over a dozen different handwriting styles.
   

   View Slide

5. 01 02 03
   Generator Discriminator ACGAN
   System Model & Methods
   Layers:
   ● Input, dense, reshape
   ● Conv2DTranspose x4
   ○ BatchNormalization
   ○ ReLU
   ● Filters: [128, 64, 32, 1]
   ● Stride: 2 (last 2 layers: 1)
   ● Kernel size: 5
   Params: 1,332,161 trainable
   (704 untrainable)
   Layers:
   ● Input
   ● Conv2D x4
   ○ LeakyReLU
   ○ alpha = 0.2
   ● Dense x3
   ● Activation layer
   ● Filters: [32, 64, 128, 256]
   ● Stride: 2 (last Conv: 1)
   ● Kernel size: 5
   Params: 1,605,638 trainable
   (0 untrainable)
   Training:
   ● batch_size = 64
   ● latent_size = 100
   ● learning_rate = 2e-4
   ● decay = 6e-8
   ● RMSprop discriminator
   ● Loss functions:
   ○ Generator images:
   binary_crossentropy
   ○ Discriminator predictions:
   categorical_crossentropy
   ● Steps = 30,000
   

   View Slide

6. Step 15000 Step 20000 Step 25000 Step 30000
   Step 500 Step 2500 Step 5000
   Results
   Step 10000
   

   View Slide

7. Discussion & Conclusion
   ● As the figures indicate, generated images became progressively more accurate over time.
   However, some graininess and noise can also be seen in some of the later images. This is not
   an error in the training of the GAN, but a representation of the dataset. Because ancient
   handwritten Manchu texts were scanned and cropped to create the training dataset, there
   was some noise present in some of the image backgrounds.
   ● Therefore, in the creation of later Manchu datasets, a better image thresholding algorithm
   will be used to minimize the background noise. Nonetheless, the results of this ACGAN research
   have demonstrated the effectiveness of possibly supplementing small-scale Manchu datasets
   with generated images to bolster neural network training.
   

   View Slide

8. 1. G. Y. Zhang, J. J. Li, A. X. Wang, “A New Recognition Method for the Handwritten Manchu Character Unit,” in
   Proceedings of the Fifth International Conference on Machine Learning and Cybernetics, Dalian, China 2006, pp.
   3339-3344, DOI: 10.1109/ICMLC.2006.258471.
   2. A. Snowberger, C.H. Lee, “A New Segmentation and Extraction Method for Manchu Character Units,” in
   Proceedings for 2022 International Conference on Future Information and Communication Engineering, Jeju, South
   Korea, pp. 42-47, 2022.
   3. R. Zheng, M. Li, J. He, J. Bi, and B. Wu, "Segmentation-Free Multi-Font Printed Manchu Word Recognition Using
   Deep Convolutional Features and Data Augmentation," in 2018 11th International Congress on Image and Signal
   Processing, BioMedical Engineering and Informatics (CISP-BMEI), Beijing, China, 2018, pp. 1-6, DOI:
   10.1109/CISP-BMEI.2018.8633208.
   4. A. Snowberger, C.H. Lee, “A Simple MNIST Style Dataset and CNN Training for Manchu Script Characters,” in
   Proceedings of the 15th International Conference on Future Information & Communication Engineering, Jeju, South
   Korea 2023, p. 136-138.
   5. R. Atienza, Advanced Deep Learning with TensorFlow 2 and Keras, 2nd ed., Packt Publishing Ltd., Birmingham, Feb.
   2020. [Online] Available: https://www.packtpub.com/book/programming/9781838821654/.
   References
   

   View Slide