Image-based Detection of Surface Defects in Concrete during Construction

TL;DR

The paper tackles automatic detection of honeycombs in concrete to reduce inspection costs by comparing two deep-learning paradigms—Mask R-CNN for instance segmentation and EfficientNet-B0 for patch-based classification—on two datasets (real-world Metis images and web-scraped images). It systematically studies data origin, labeling strategies, and transfer-learning regimes, revealing that web data alone may not generalize to real construction sites and that dataset diversity is crucial. EfficientNet-B0 achieves state-of-the-art-like performance on related benchmarks, while Grad-CAM analyses provide insight into model focus and potential biases. The work contributes HiCIS and HiCC datasets for ongoing research and discusses practical considerations, including when to use segmentation versus patch classification and how active learning could enable scalable defect documentation systems in practice.

Abstract

Defects increase the cost and duration of construction projects as they require significant inspection and documentation efforts. Automating defect detection could significantly reduce these efforts. This work focuses on detecting honeycombs, a substantial defect in concrete structures that may affect structural integrity. We compared honeycomb images scraped from the web with images obtained from real construction inspections. We found that web images do not capture the complete variance found in real-case scenarios and that there is still a lack of data in this domain. Our dataset is therefore freely available for further research. A Mask R-CNN and EfficientNet-B0 were trained for honeycomb detection. The Mask R-CNN model allows detecting honeycombs based on instance segmentation, whereas the EfficientNet-B0 model allows a patch-based classification. Our experiments demonstrate that both approaches are suitable for solving and automating honeycomb detection. In the future, this solution can be incorporated into defect documentation systems.

Figures (10)

• Figure 1: Grad-CAM of EfficientNet-B0 trained on different datasets for an image containing pebbles from metis-s224-p224/test
• Figure 2: $(a)$ and $(b)$ show example images with patch-wise classification. $(c)$ and $(d)$ show the corresponding activations.
• Figure 3: False positives for non-concrete patches by our finetuned EfficientNet-B0 trained on different training sets
• Figure 4: Untypical scaling of a honeycomb by our finetuned EfficientNet-B0 trained on different training sets
• Figure 5: Mask R-CNN training