Abstract
Deep Neural Networks (DNNs) have become essential tools for Structural Health Monitoring (SHM), particularly in crack detection for bridges and roads. However, these models often suffer from miscalibration, manifesting as both over-confidence in incorrect predictions and under-confidence in correct ones. This paper introduces Attack-inspired Calibration Loss (AICL), a novel framework that utilizes adversarial robustness as a proxy for prediction correctness. By testing the stability of feature embeddings under attack, AICL effectively regularizes model confidence, leading to significant improvements in reliability across multiple benchmarks without compromising predictive accuracy.
Introduction
In safety-critical applications like SHM, a model's confidence is as vital as its accuracy. A miscalibrated model might classify a structural crack as "no crack" with high certainty, potentially leading to catastrophic failures if inspections are skipped. Crack recognition presents a unique challenge: models are often under-confident due to low signal-to-noise ratios and over-confident when mistaking noisy backgrounds for cracks. Existing calibration techniques designed for general image sets often fail to address this "double-sided" miscalibration issue.
Methodology
The proposed AICL framework consists of two primary components:
- Attack-inspired Correctness Estimation (ACE): This module evaluates the "honesty" of a prediction by measuring its robustness against adversarial attacks (specifically Projected Gradient Descent, or PGD). Instead of attacking raw pixels, which is computationally expensive, ACE targets feature embeddings. The number of attack rounds required to flip a prediction serves as an indicator of correctness: fragile predictions flip quickly, while robust ones are likely correct.
- Correctness-aware Distribution Guidance: Rather than forcing individual samples to match specific confidence scores, AICL employs a distribution-level approach. It groups samples based on their robustness (from the ACE step) and ensures that, on average, robust groups maintain higher confidence than fragile ones. This "plug-and-play" loss function can be integrated into standard architectures like ResNet or VGG.
Experimental Results
The method was evaluated on three benchmarks: SDNET2018, CRACK2019, and a newly introduced Bridge2024 dataset containing high-resolution drone imagery of real-world structural challenges.
Key Findings
- Superior Calibration: AICL consistently outperformed state-of-the-art methods such as MbLS and DECE. On the Bridge2024 dataset, the Expected Calibration Error (ECE) was reduced from a baseline of 6.87% to 2.10%.
- Maintained Performance: Unlike some calibration methods that trade off accuracy for reliability, AICL often improved the F1-score by helping the model better distinguish between actual cracks and environmental noise.
- Robustness as Truth: The core insight of the study is that adversarial robustness is a more reliable indicator of truth than raw softmax probabilities in noisy SHM environments.
Conclusion and Future Work
AICL provides a robust solution for making crack recognition models more trustworthy for real-world deployment. Future research will focus on reducing the computational overhead of adversarial training and integrating these calibrated models into autonomous robotic systems for real-time structural inspections.
Summary updated