A Comparative Analysis of Two Deep Learning Neural Networks for Defect Detection in Steel Structures Using UAS

Abstract

Steel is used in different infrastructural constructions. The durability and serviceability of steel made it more suitable than other construction materials. However, exposure to weather elements can cause defects in steel structures. Early detection and treatment of structural defects can prevent the structure from becoming more damaged and more expensive to repair. Corrosion resistance and fatigue strength in any steel structure can be influenced by defects such as patches, scratches, and coating erosion. Current methods to detect steel defects are based on manual visual inspection. Autonomous UAS imaging-based defect detection methods have shown promising results in terms of accuracy and time. This paper compares the performance of two deep learning models, InceptionResnetV2 and ResNet152V2, for detecting steel defects. These models were trained in transfer learning mode and tested on two different datasets, the Severstal dataset present on Kaggle and a dataset generated by the authors of this paper. The results show that ResNet152V2 outperforms InceptionResnetV2 with an average accuracy of 95% and a misdetection rate of 5%. Overall, both the models, ResNet152V2 and InceptionResNetV2, showed an improvement of 12.59% and 9.59%, respectively, compared to MobileNet used in a previous study, when all were trained and tested on the Severstal dataset.