Speaker
Description
Infrared thermography (IRT) is a well-established non-destructive testing (NDT) technique widely used for detecting and characterizing defects in materials and structures. It offers a contactless and wide-area inspection capability, enabling the rapid evaluation of thermal responses that indicate subsurface anomalies. However, in many cases, raw thermographic data do not exhibit clear defect patterns due to noise, emissivity variations, or environmental interferences. Consequently, post-processing techniques have become essential to enhance defect visibility and extract reliable diagnostic information.
Among the most recognized approaches are statistical moment analysis, Pulse Phase Thermography (PPT), Principal Component Thermography (PCT), and Thermographic Signal Reconstruction (TSR). These methods respectively aim to highlight subtle thermal variations, separate defect-related information in the frequency domain, reduce data dimensionality to isolate dominant thermal patterns, and smooth temperature evolution curves to improve contrast. While these methods have demonstrated significant potential in controlled environments, their deployment in industrial settings remains limited by computational constraints. Processing high-resolution thermographic sequences with thousands of frames requires substantial computational power, which challenges real-time implementation in in-line inspection systems.
This work investigates the acceleration of classical post-processing methods for infrared thermography through the use of modern hardware platforms, particularly Graphics Processing Units (GPUs) and edge computing devices. Optimized implementations of several established algorithms are presented, leveraging efficient vectorized operations and the parallel computation capabilities of GPUs. The proposed implementations achieve remarkable performance gains, reducing execution times to below 35 ms for sequences of 2000 frames of 512 × 512 pixels—equivalent to over 524 megapixels processed per sequence. These results demonstrate that GPU- and edge-based architectures can enable real-time or near-real-time defect analysis, significantly improving the practical feasibility of IRT in demanding industrial environments.
This work also provides a detailed analysis of the computational optimization strategies that contribute to the achieved acceleration. Techniques such as vector data processing are exploited to maximize throughput. A comparative study between different GPU architectures and embedded systems demonstrates that even low-power edge devices can deliver competitive performance when algorithms are carefully tuned to their hardware characteristics. This finding is particularly relevant for applications where portability, energy efficiency, and cost are major constraints.
Beyond acceleration, the study explores the scalability and deployability of these techniques on embedded edge computing platforms. The experiments confirm that edge devices can perform most post-processing operations locally, reducing data transmission demands and allowing for low-latency decision-making in field conditions. This capability is essential for autonomous inspection systems, predictive maintenance, and continuous quality monitoring aligned with Industry 5.0 principles, which emphasize human-centricity, resilience, and sustainability.
The implementations developed in this work are released as open-source software to promote transparency, reproducibility, and adoption within the NDT community. The availability of these tools is expected to facilitate benchmarking, encourage collaborative research, and support the development of next-generation thermographic inspection systems. By bridging the gap between algorithmic development and practical deployment, this study demonstrates that optimized GPU and edge computing solutions can transform infrared thermography from a laboratory analysis technique into a powerful real-time industrial diagnostic tool.
