Lightweight AI spots hidden concrete defects from infrared images

Engineers have long struggled to find damage they cannot see. Beneath the surface of a concrete bridge deck or pier, cracks, voids and delaminations can quietly grow until they trigger spalling that threatens structural integrity. Infrared thermography offers a way to peer beneath the surface without contact, but interpreting the resulting thermal images has remained a manual, expertise-dependent task prone to human error. A new study now presents a compact deep learning model, called CA-YOLO-IRT, that automates this interpretation while running fast enough for real-time inspection.

The work, published in Case Studies in Construction Materials, was carried out by Quang Tai Ta, Jungwon Huh, Van Ha Mac, Ba Huu Dinh and In-Tae Kim. Their starting point was a practical observation: infrared thermography is among the most efficient nondestructive testing methods for subsurface defects invisible to the naked eye, yet the analysis of thermographic images still relies heavily on the investigator’s skill. Combining thermography with deep learning promises to cut inspection time and cost while reducing the need for specialized expertise, and the team set out to design a detector tailored specifically to the unusual character of thermal imagery.

Subsurface defects show up in infrared images in a distinctive way. Because an air-filled void or crack conducts heat far more poorly than concrete, roughly 0.024 watts per meter-degree Celsius compared with about 1 watt per meter-degree Celsius, it disrupts heat flow when the structure is warmed by the sun or by artificial lamps. The disrupted heat transfer produces localized temperature differences on the surface, which an infrared camera records as abnormal thermal patterns against the surrounding intact material. But these signatures are typically weak in contrast, indistinct at their boundaries and almost devoid of texture, unlike the rich, high-frequency edge patterns in ordinary photographs that most modern detectors were designed to exploit.

That mismatch motivated a task-specific redesign rather than an off-the-shelf model. The researchers built on YOLOv5n, a lightweight object detector their earlier work had identified as offering a favorable balance between detection accuracy and computational efficiency for concrete delamination detection. Object detection, which outputs defect classes and bounding boxes, was chosen over pixel-level segmentation because it provides a practical trade-off between localization and deployment feasibility for field inspection. In the new architecture, the network’s width multiplier was halved from 0.25 to 0.125, and every C3 convolutional block in the backbone was replaced with a coordinate attention module, or CA, which embeds positional information into channel attention by pooling features separately along horizontal and vertical directions.

Several design choices were tuned specifically for thermal data. The CA reduction ratio, which controls how much the intermediate attention representation is compressed, was set to 1 instead of the conventional value of 32, avoiding an information bottleneck that could discard the subtle thermal variations critical for defect localization. Attention was confined to the backbone, leaving the neck and detection head unchanged to preserve stable multi-scale feature fusion. The resulting model contains just 388,446 parameters and requires 0.9 gigafloating-point operations, a fraction of the 1,760,518 parameters and 4.1 gigafloating-point operations of the original YOLOv5n baseline.

The evidence for these choices came from a systematic series of experiments. The team trained all models from scratch on a self-prepared dataset of 10,000 infrared images from twelve concrete specimens fabricated with artificial defects, captured with a FLIR SC660 long-wave infrared camera under both active thermography with halogen heating and passive solar-loaded thermography over 24-hour periods. An additional 2,000 images from two unseen specimens served as an independent validation set, with training and evaluation data kept strictly separated to prevent cross-specimen leakage. Progressive tests showed that simply adding attention modules after the convolutional blocks increased complexity without improving performance, whereas replacing the blocks outright with coordinate attention delivered the best results.

The ablation studies also revealed that more attention is not always better. Extending coordinate attention into the neck of the network, where multi-scale features are fused, raised the parameter count and cut the key AP0.5:0.95 metric by 1.8 percentage points on laboratory data, a gap that widened dramatically in the field. Comparisons against alternative attention mechanisms told a similar story: a CBAM-based variant edged out the proposed model on laboratory validation, but its advantage evaporated on real-world images, while the coordinate attention design held up better. Setting the reduction ratio to 1 likewise proved its worth mainly under real-world conditions rather than in the laboratory.

The decisive test came on 105 thermal images collected from real concrete structures, including bridge decks, beams and piers in Gwangju, South Korea, photographed on sunny days between 10:00 and 16:00. Here the proposed model achieved 70.4 percent precision, 48.3 percent recall, 54.5 percent AP0.5 and 25.8 percent AP0.5:0.95, the best precision and average precision among all candidates. Benchmarked against newer lightweight detectors including YOLOv8n, YOLO11n, YOLOv13n and YOLO26n, the compact model improved precision by 8.3 percentage points and AP0.5 by 6.0 percentage points over the YOLOv5n baseline while using roughly 78 to 89 percent fewer parameters and computations than its competitors, and it sustained an inference speed of 44 frames per second on an RTX 3090, above the commonly adopted 30-frame threshold for real-time processing.

Interpretability analysis with the Eigen-CAM visualization technique supported the design rationale. Heat maps showed that the original YOLOv5n scattered its activations broadly across thermal images, often focusing strongly on irrelevant background regions, whereas the proposed model concentrated its responses tightly around the delamination areas while suppressing background noise. The researchers interpret this as evidence that coordinate-aware spatial encoding, rather than generic texture-based feature extraction, is the key mechanism allowing the network to activate discriminative thermal regions in low-texture infrared images.

The authors are candid about the limits of the work. The substantial drop in performance from laboratory to field conditions, with AP0.5 falling from 99.2 percent to 54.5 percent for the proposed model, reflects a significant domain gap driven by weaker thermal signals from deeper defects, uneven solar heating, wind effects, emissivity variations and background thermal interference, and field annotations themselves were only partially confirmed by hammer sounding. The laboratory dataset included defects only up to 5 centimeters deep, the model handles a single defect class, and edge-device deployment has yet to be verified. Even so, the team argues that the architecture offers a practical foundation for automated thermographic inspection, and that future training on field-collected images of naturally occurring defects should narrow the gap between the laboratory and the bridge deck.

Subject of Research: A lightweight deep learning model for automated detection of subsurface defects in concrete using infrared thermography

Article Title: CA-YOLO-IRT: A lightweight thermographic detector for automated detection of subsurface defects in concrete structures using infrared thermography

Article References: Ta, Q. T., Huh, J., Mac, V. H., Dinh, B. H., & Kim, I.-T. (2026). CA-YOLO-IRT: A lightweight thermographic detector for automated detection of subsurface defects in concrete structures using infrared thermography. Case Studies in Construction Materials, 25, Article e06623. https://doi.org/10.1016/j.cscm.2026.e06623

Image Credits: AI Generated

DOI: 10.1016/j.cscm.2026.e06623

Keywords: infrared thermography, concrete defects, deep learning, object detection, structural health monitoring, delamination, YOLOv5n, coordinate attention, nondestructive testing, lightweight models, bridge inspection, thermal imaging