Year: 2025
Type: conference-paper
Abstract: One common data augmentation method used in text image recognition is color shift transformations. A widely used method is the Variational Color Shift (VCS), which can automatically predict a range of color shifts for each image. The algorithm has significant overfitting when applied to text images that are highly interfered with and contain noise, which hinders further improvements in text recognition accuracy. In this study, we submit a new transformed layer inside the VCS framework using cyclic boundary shifts over the text images. A new module, Transformed Variational Color Shift (TVCS) is proposed. This method integrates a generalized set of colors (the original dataset) in a manner that can transform the input dataset. The experimental results show that TVCS has superior recognition accuracy than VCS and has better robustness against interference. The algorithm proposed is innovative as the data augmentation of text images, especially in a noisy environment.
Authors: Xing Wan, Juliana Johari, Fazlina Ahmat Ruslan
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4411014986