A Latent Diffusion Based Image Generation Method for Anomaly Detection

The primary issue in visual anomaly detection is lack of sufficient samples, which makes it difficult for real-world applications unless an effective anomalous image generation algorithm is proposed. To this end, this paper formulates a latent diffusion model to generate a variety of anomalous images by adapting the initial latent tensors to the anomaly. Therefore, the diffusion process is speeded up via transferring the image synthesis in a low-dimensional space, with anomalous labels provided. The pseudo-anomaly images and pseudo-labels are then used for data augmentation for anomaly detection based on YOLO-v8, and quantitative and qualitative experiments on NEU-DET demonstrate the applicability and superiority of our approach on precision and diversity.