Dataset: Dataset and Code for RACON Thermographic Anomaly Detection

Abstract

A lightweight, training-free framework for thermographic anomaly detection based on Random Convolution Kernels called RACON. Unlike conventional deep learning approaches that require extensive labelled datasets and heavy models, our method operates in a zero-shot setting , using randomly generated filters combined with multi-scale Gaussian pooling, group convolution, and feature-space anomaly scoring. Experiments on a CFRP composite thermographic dataset demonstrate that the proposed method achieves 94.0% I-AUROC and 92.19% P-AUROC, outperforming PatchCore and matching U-Net performance, while using 10× less memory and 3× lower latency. These results validate RACON as an efficient alternative for real-time industrial inspection on resource-constrained devices.