Concrete Bridge Defects Identification and Localization Based on Classification Deep Convolutional Neural Networks and Transfer Learning

Conventional practices of bridge visual inspection present several limitations, including a tedious process analyzing images manually to identify potential damages. Vision-based techniques, particularly Deep Convolutional Neural Networks, have been widely investigated automatically identify, localize, and quantify defects in images. However, massive datasets with different annotation levels are required train these deep models. This paper presents dataset more than 6900 featuring three common concrete bridges (i.e., cracks, efflorescence, spalling). To overcome the challenge limited training samples, Transfer Learning approaches fine-tuning state-of-the-art Visual Geometry Group network were studied compared classify defects. The best-proposed approach achieved high testing accuracy (97.13%), combined F1-scores 97.38%, 95.01%, 97.35% for spalling, respectively. Furthermore, effectiveness interpretable networks was explored context weakly supervised semantic segmentation using image-level annotations. Two gradient-based backpropagation interpretation techniques used generate pixel-level heatmaps localize test Qualitative results showcase use maps provide relevant information on defect localization weak supervision framework.