An Evaluation of Diffusion-Based Anomaly Detection in Metal Can Products

Deep learning has been considered a successful solution to process images and complex data. Hence, it is expected to be a standard solution for quality assurance in manufacturing. A standard deep learning-based method tries to reconstruct a normal image, and the difference between a testing image and its corresponding reconstructed image serves as an anomaly map. While a reconstruction model had been proposed, diffusion methods had been considered as SoTA solutions for image generation. However, these methods are tested on well-prepared datasets collected by costly devices in less noisy conditions. In an industrial environment, the image may be collected using economic hardware with serious noise. Motivated by the observation, this work tries to evaluate how diffusion-based anomaly detections work in a practice environment. We first collected a dataset by ourselves and tested it on various well-known anomaly detection methods. The hardware includes a rotating disk and a camera to capture sample data from various angles, while the software serves as an anomaly detection system for test samples. Also, we focus on well-known diffusion models to address whether this method works in high-variance environments. Comprehensive experiments on three can-datasets had been implemented, and the result shows that at the image level, the diffusion method works robustly without error.