[4I2-J-2] Machine learning: uncertainty of targets

[4I2-J-2-03] Deep Unsupervised Anomaly Segmentation via Aleatoric Uncertainty-Aware Score

〇Kazuki Sato¹, Kenta Hama¹, Takashi Matsubara¹, Kuniaki Uehara¹ (1. Kobe University)

Keywords:Deep Learning, Unsupervised Learning, Anomaly Detection

Image-based anomaly segmentation is a basic topic in the field of image analysis. Especially, unsupervised training is preferred when dealing with unknown types of anomalies. For this purpose, probabilistic models trained to maximize the likelihood of known samples are employed to identify samples with low estimated likelihoods as anomalies. However, they tend to be sensitive to complex structures rather than semantic anomalies. In this paper, we propose a novel uncertainty-aware score for anomaly segmentation by removing the term that reflects the data complexity from the approximated log-likelihood. Experimental results demonstrate the robustness of the proposed score with respect to data complexity.