Projection Regret: Research on Reducing Background Bias and Novelty Detection via Diffusion Models

To address the data imbalance problem that arises in manufacturing processes and other settings, Novelty Detection technology—which learns only from normal data to identify outliers—is important. Existing generative model-based approaches had limitations in that they failed to properly distinguish outliers whose backgrounds resemble normal data, or suffered from overfitting to background information (Texture Overfitting).

To solve this, LG AI Research proposed a Projection Regret methodology utilizing a Consistency Model. This approach involves diffusing (Forward Diffusion) the outlier data at an appropriate timestep and then restoring it (Reverse Diffusion). This takes advantage of the characteristic that the background is preserved while only the semantic information is modified to align with normal data.

Since simply measuring semantic distance through LPIPS alone can result in background information becoming mixed in, a new metric was designed that calculates the distance difference between the additionally edited image and the transformed image that includes background information. This demonstrated that background bias can be reduced and outlier detection performance can be effectively improved.