Object anomaly detection using infrared images

Infrared (IR) imaging plays a critical role in modern industrial inspection systems, enabling the detection of subsurface defects such as delaminations, cracks, and thermal inconsistencies that are often invisible in visible-spectrum images. However, developing accurate and robust IR-based anomaly detection models remains challenging due to the low texture of thermal images, smooth intensity gradients, and limited annotated data. In this work, we propose a deep learning framework for IR anomaly detection that learns only from defect-free images. Our method begins with multi-scale feature extraction and incorporates a High-Frequency Gradient Fusion (HFGF) module to inject directional edge cues into the features. This is followed by a two-stage attention mechanism: Directional Saliency Attention (DSA) highlights orientation-aligned structures, and Multi-Scale Attention Refinement (MSAR) gathers contextual information at multiple scales. To simulate realistic anomalies without altering input images, we introduce Saliency-Aware Feature Perturbation (SAFP), which injects structured noise into salient feature regions. A lightweight discriminator learns to separate normal and perturbed features, enabling effective anomaly detection. Importantly, the four modules operate synergistically: HFGF supplies edge cues that steer DSA and inform SAFP, while MSAR preserves and propagates those orientation-aware signals across scales, forming a closed loop that enhances IR feature representations. The framework is compact and designed for practical scalability, making it adaptable to different IR sensors and industrial scenarios such as photovoltaic panel maintenance and industrial part inspection with modest calibration or lightweight fine-tuning. On public datasets, our method surpasses the state of the art by +7.7% for anomaly detection and +5.7% for anomaly localization, demonstrating strong potential for practical IR-based inspection. • An unsupervised deep learning framework for fast and accurate infrared anomaly detection, surpassing the state of the art by +7.7% in anomaly detection and +5.7% in anomaly localization. • High-Frequency Gradient Fusion (HFGF) enriches pretrained infrared features with directional gradients, enhancing sensitivity to fine thermal transitions. • Directional Saliency Attention (DSA) highlights orientation-aligned and low-frequency features, learning spatial priors from normal data. • Multi-Scale Attention Refinement (MSAR) adaptively refines features across scales using orientation-aware routing and channel-wise recalibration to model diverse thermal structures. • Saliency-Aware Feature Perturbation (SAFP) injects structured feature variations to simulate pseudo-anomalies and improve training under normal-only supervision.