LG AI Research 490

In fields with limited data, such as manufacturing, detecting defects using only normal data is very difficult. Existing anomaly detection methods based on Generative Models had the limitation of overfitting to image complexity or background information (Texture Overfitting), making it difficult to properly identify abnormal data whose background resembles normal data.

LG AI Research proposed Projection Regret research using a Consistency Model to solve this problem. This approach leverages the characteristic that when an image is edited through the Diffusion process, the background is preserved while only the semantic information is restored to normal data.

The specific core mechanism is as follows.

• Using Semantic Distance: Measures the difference between the original image and the Diffusion-based edited image.
• Improved LPIPS Metric: Instead of simple LPIPS, an additional Editing step is performed to remove the portion occupied by background information, and the difference between the two Distance Outputs is used as the final Metric.

This technology has demonstrated superior performance compared to existing Diffusion-based algorithms and Unsupervised-based methods across various image domains.

This article also introduces additional research that improves the performance of Diffusion models.

• Google DeepMind Research: Designed an optimal Weighting configuration for combining the Training Loss of Diffusion models, leading to performance improvements on the ImageNet dataset.
• Zero-Shot Classifier: Covers research showing that Text-to-Image Diffusion models can serve as classifiers.