Image quality enhancement using HybridGAN for automated railway track defect recognition

Numerous researchers have explored the potential of artificial intelligence to improve the rail inspection process, which required high quality in terms of resolution and clarity images. However, poor resolution and blurriness are chronic, difficult-to-resolve issues in dynamically acquired railway inspection images. Therefore, a novel HybridGAN is proposed that combines ESRGAN and DeblurGANv2 to improve the quality of dynamically acquired images in terms of resolution and blurriness. To validate the performance, results obtained using a YOLOv4 model trained and tested, respectively, on the original dataset and on an improved dataset were compared. Furthermore, three sensitivity analyses were conducted to test the performance. HybridGAN consistently demonstrated the ability to improve mAP scores that were 8.76-10.83% better across multiple resolution levels and 18.80-28.07% higher on the low-quality dataset. HybridGAN offers an effective solution to the problem of using low-quality image datasets both low resolution and high blurriness in automated railway safety maintenance.

Automated summary

This study proposes a novel HybridGAN that combines ESRGAN and DeblurGANv2 to improve the resolution and blurriness of dynamically acquired railway inspection images. Experimental results show that HybridGAN consistently improves mAP scores across multiple resolution levels and performs significantly better on low-quality dataset.