Speaker
Description
Active infrared thermography (IRT) is a contactless and non-intrusive technique for non-destructive evaluation, providing a fast means for locating and sizing subsurface defects in a wide range of materials. The virtual wave concept (VWC) is a prominent approach for performing thermographic reconstruction, which leverages ultrasonic methods for accurately inferring the internal structures of objects under evaluation. However, performing the VWC reconstruction process from surface temperature measurements is challenging, as it requires solving a series of ill-posed inverse problems. Although deep learning techniques have recently shown promising results in solving the underlying inverse problems, the evaluation and optimization of such approaches remains a critical challenge in IRT, as commonly used pixel-level metrics do not adequately represent practical reconstruction quality. To address this issue, we introduce an object-level evaluation framework aligned with the practical goal of industrial inspection: detecting individual structural faults rather than achieving pixel-perfect reconstruction. Our results on large-scale simulated datasets demonstrate the effectiveness of this approach in providing an accurate and detailed quantitative evaluation of DL-based reconstruction techniques. Additionally, the presented experiments highlight the alignment between qualitative evaluation and the proposed quantitative approach.
