Semantic Segmentation of Defects in Infrared Thermographic Images of Highly Damaged Concrete Structures

There is a global research trend to enhance condition assessment of the concrete infrastructure by the development of advanced nondestructive testing (NDT) methods. Computer vision–based systems have been developed to detect different types of defects in both regular and thermographic images because these systems could offer a timely and cost-effective solution and are able to tackle the inconsistency issues of manual assessment. This paper investigates the performance of different deep neural network models to detect main concrete anomalies, including delamination, cracks, spalling, and patches in thermographic and regular images captured from a variety of distances and viewpoints. These models were trained and tested using images taken from a century-old buttress dam and validated in images captured from the decks of two concrete bridges. The results showed that the MobileNetV2 had promising performance in the identification of multiclass damages in the thermal images, identifying 79.7% of the total delamination, cracks, spalling, and patches on the test images of highly damaged concrete areas. The VGG 16 model showed better precision by reducing the number of false detections.