PRECISION COMPARISON OF DEEP LEARNING MODELS FOR DETECTING CONCRETE SURFACE DETERIORATION TYPES FROM DIGITAL IMAGES

Abstract

 Visual inspection of concrete structures is an important task; however, it is labor-intensive work, taking much time and effort, and it is often dangerous when inspectors work in high places. Automatic detection of deterioration such as cracks, free lime, exposed reinforcing bars, etc., from digital images has been extensively researched. Current deep learning approaches have improved the detection performance compared with conventional approaches. Most deep learning approaches use publicly available deep learning models, such as Faster R-CNN and Single Shot multiBox Detector (SSD). Although the results of deterioration detection methods have been compared with the conventional approach, the deep learning models themselves have not been compared. In this research, using the same datasets, seven deep learning models (YOLOv3, RetinaNet-50, RetinaNet-101, RetinaNet-152, SSD512, SSD300, and Faster R-CNN), were compared for detecting five types of deterioration (cracks, exposed reinforcing bars, free lime c-type, free lime d-type, and free lime e-type). YOLOv3 showed the highest mean average precision (mAP) of 85.7%, whereas the other models showed less than 80%.