Comparative Effectiveness of Data Augmentation Using Traditional Approaches versus StyleGANs in Automated Sewer Defect Detection

Abstract

This study compared how traditional data augmentation and the state-of-the-art style-based generative adversarial network (StyleGAN) benefit automated sewer defect detection using a You Only Look Once (YOLO) object detection model. There were 25 augmentation scenarios (13 individual augmentations, 8 combined augmentations, and 3 StyleGAN augmentations, as well as a baseline) investigated, with 3 types of data set sizes (300, 600, and 1,200 images) and 4 types of sewer defects (disjoint, obstacle, residential wall, and tree root). Results showed that geometric transformations generally outperformed color transformations. Among the traditional methods, random cropping worked best to improve the detector performance, meaning that combined effects may not be as strong as a single type of augmentation. It was also noted that not all augmentation measures benefited the detection, as in our case, 24% of the augmentation effects resulted in lower detection accuracy than the baseline scenario. StyleGAN performed remarkably well in improving the data quality through increasing style and diversity. Moreover, data augmentation had a greater impact on improving the detection of residential wall and tree root [with an increase in the mean of average precision (AP) of 32% and 22%] in comparison to disjoint and obstacle (21% and 16%), respectively. The findings will benefit the future selection and use of augmentation methods in enhancing the performance of deep learning models.