Self-supervised image anomaly detection with distortion perception and rectification

Lens distortion in industrial imaging is a critical, yet overlooked, problem that causes high false-positive rates in anomaly detection systems. Current methods lack the adaptability to handle these common spatial variations. To solve this, we introduce a novel approach that explicitly learns to model these distortions. We propose a differentiable distortion field learning module guided by a specialized distortion learning loss function. Our method disentangles true defects from imaging artifacts, significantly improving the robustness and accuracy of anomaly detection in real-world settings. Our method is built on a Teacher-Student framework. The student network first rectifies an input image using a differentiable distortion correction module. To ensure geometric accuracy, the predicted correction field is directly supervised by comparing it to a ground-truth transformation grid. Subsequently, we minimize the first-order reconstruction error between the features of the teacher and the corrected student. This objective directly enhances the quality of the distortion correction by enforcing featurelevel alignment. We embedded our distortion correction module into the EfficientAD model and evaluated its anomaly detection performance on both original and geometrically distorted images. The experimental results demonstrate that our proposed method generally outperforms the baseline EfficientAD, achieving higher overall AUROC and F1-Score metrics.