Rust Detection on Galvanized Iron Sheets using Convolutional Neural Network and Mask R-CNN

Rust in galvanized iron sheets poses problems with the structural quality of the sheet. It is important that regular maintenance and monitoring are employed to prevent further deterioration and hazards. Smart methods have emerged in structural maintenance and monitoring through computer technology. This study aimed to classify rust based on visibility using Convolutional Neural Network (CNN) and locate rust using Mask Region-based Convolutional Neural Network (Mask R-CNN). To do this, a manually acquired and custom rust dataset was built and categorized into No Rust, Slightly Visible Rust, Visible Rust, and Heavy Visible Rust. The custom dataset was built with the guidance of an expert on the subject matter and was validated and verified. In this study, CNN obtained 80% accuracy on the test set. Mask R-CNN obtained an Average Precision with a threshold of 50% (AP50) value of 70.958% on its box parameter and an AP50 value of 70.139% on its segment parameter. The prototype was also tested under different lighting intensities, and it was found that the classification and localization performance had improved and became more consistent when light intensities reached 160,000 lux.