Enhancing random surface anomaly detection in real-world using a four-stage one-class approach

Defect detection and localization are critical for quality control in manufacturing, yet existing algorithms and models trained on laboratory datasets often fail in real industrial scenarios due to their static nature, especially in non-mass production. Moreover, limited and heterogeneous defective samples, coupled with costly human annotation, highlight the need for unsupervised methods relying solely on normal images. To address these challenges, we propose the Random Surface Anomaly Detection (RSAD) model, a four-stage one-class anomaly detection and localization approach. Initially, leveraging embedding-based techniques, we introduce transfer learning with a pretrained ImageNet network in extracting locally aggregated features. Next, adapter tuning is applied to transfer these features into the industrial domain, reducing bias towards natural images. Additionally, random Gaussian noise is introduced into normal feature representations within the feature space and a discriminator then scores feature normality. Finally, experiments on the MPDD dataset and other benchmarks, demonstrate the RSAD model's state-of-the-art (SOTA) performance in anomaly detection, validating its trustworthiness in real-world manufacturing environments.