AUTHOR=Wang Shiyong , Khan Asad , Lin Ying , Jiang Zhuo , Tang Hao , Alomar Suliman Yousef , Sanaullah Muhammad , Bhatti Uzair Aslam 

TITLE=Deep reinforcement learning enables adaptive-image augmentation for automated optical inspection of plant rust

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 14 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2023.1142957

DOI=10.3389/fpls.2023.1142957

ISSN=1664-462X

ABSTRACT=Deep learning-based automated optical inspection can benefit from image augmentation which enlarges the image quantity for training and testing. However, one significant challenge is that any single image augmentation method cannot achieve consistent performance over all the images. To address this issue, the specific DRL algorithm, DQN, is used in this study to organize an adaptive image augmentation scheme. The DQN is assisted with the geometric and pixel indicators for state extraction, the DeepLab-v3+ model for verifying the augmented images and generating the reward, and the image augmentation methods as actions. The proposed DRL-enabled adaptive image augmentation framework achieves better augmentation performance than any single image augmentation method and better segmentation performance than the other semantic segmentation models. The experimental results confirm that the DRL-enabled adaptive image augmentation framework can adaptively select augmentation methods that best match the images and the semantic segmentation model.