Training-free thermographic anomaly detection using random convolution kernels

Abstract

Thermographic inspection has been widely used for non-destructive evaluation of composite structures, particularly for detecting impact-induced defects in carbon fibre reinforced polymer (CFRP) materials. However, most existing approaches rely on supervised learning models that require large amounts of labeled data and significant computational resources, limiting their applicability in practical industrial environments. In this work, a lightweight and training-free framework based on random convolution features is proposed for thermographic inspection. The method eliminates the need for model training and labeled datasets by employing randomly generated convolutional kernels to extract discriminative thermal patterns, followed by simple statistical decision mechanisms for both defect classification and pixel-wise segmentation. This enables rapid deployment and efficient processing without compromising interpretability. Experimental results on CFRP impact damage data demonstrate that the proposed approach achieves competitive performance compared to both traditional feature-based methods and deep learning models. In classification tasks, the method consistently outperforms conventional handcrafted features, while in segmentation tasks it achieves a pixel accuracy of 99.5%, comparable to the best-performing baseline. Importantly, the proposed framework significantly reduces computational cost, with negligible model size and low memory consumption compared to data-driven approaches. Overall, the results indicate that the proposed method provides a favorable balance between accuracy and efficiency, making it particularly suitable for real-time and resource-constrained thermographic inspection scenarios.