Small Objects and Faults Detection on Corrugated Metal Roof using Drone equipped with Deep Learning

Corrugated Metal roofs are commonly used at industrial buildings due to their ease of installation and maintenance. However, they need regular maintenance due to several environmental factors such humidity, and wind. In this study, a method for automatic fault detection on metal roofs is presented that utilizes a drone and an object detection algorithm. An image inpainting-based simple image augmentation method is proposed that can minimize the problem of class imbalance. It can be used with any object detection algorithm. Moreover, object detection algorithms incorporate negative training that uses background as negative samples. However, when negative training is applied to images with large background to foreground ratio, it leads to inefficient training. Therefore, a technique is proposed to mitigate the problem of easy negatives by narrowing down the region of interest. Finally, the results suggest that this approach can be further extended to other rooftop faults such as pipe defects, corrosion, and cracks.