[Paper Review] AnoSeg: Anomaly Segmentation Network Using Self-Supervised Learning

Jou Won Song, Kyeongbo Kong+3 more|arXiv (Cornell University)|Oct 7, 2021

Anomaly Detection Techniques and ApplicationsComputer Science24 references37 citations

TL;DR

AnoSeg directly generates pixel-wise anomaly maps using self-supervised learning with hard augmentation, adversarial reconstruction, and coordinate channels, achieving state-of-the-art anomaly segmentation and competitive detection on the MVTec AD dataset.

ABSTRACT

Anomaly segmentation, which localizes defective areas, is an important component in large-scale industrial manufacturing. However, most recent researches have focused on anomaly detection. This paper proposes a novel anomaly segmentation network (AnoSeg) that can directly generate an accurate anomaly map using self-supervised learning. For highly accurate anomaly segmentation, the proposed AnoSeg considers three novel techniques: Anomaly data generation based on hard augmentation, self-supervised learning with pixel-wise and adversarial losses, and coordinate channel concatenation. First, to generate synthetic anomaly images and reference masks for normal data, the proposed method uses hard augmentation to change the normal sample distribution. Then, the proposed AnoSeg is trained in a self-supervised learning manner from the synthetic anomaly data and normal data. Finally, the coordinate channel, which represents the pixel location information, is concatenated to an input of AnoSeg to consider the positional relationship of each pixel in the image. The estimated anomaly map can also be utilized to improve the performance of anomaly detection. Our experiments show that the proposed method outperforms the state-of-the-art anomaly detection and anomaly segmentation methods for the MVTec AD dataset. In addition, we compared the proposed method with the existing methods through the intersection over union (IoU) metric commonly used in segmentation tasks and demonstrated the superiority of our method for anomaly segmentation.

Motivation & Objective

• Motivate direct pixel-wise anomaly segmentation rather than relying on post-hoc localization or reconstruction errors.
• Develop a self-supervised framework to generate accurate anomaly maps from normal data and synthetic anomalies.
• Leverage data augmentation, adversarial learning, and positional information to improve segmentation fidelity.
• Provide an anomaly detector that benefits from the generated anomaly maps to enhance detection performance.

Proposed method

• Use hard augmentation to create synthetic anomaly data and corresponding ground-truth masks from normal data.
• Train AnoSeg in a self-supervised manner with pixel-wise segmentation loss on synthetic data.
• Incorporate an adversarial loss with a discriminator that evaluates normal-region distribution using input-image and anomaly-map pairs.
• Apply masked reconstruction loss so reconstruction focuses on normal regions, aided by a generator–discriminator setup.
• Concatenate coordinate channels (CoordConv-inspired) to provide pixel-position information to the network.

Experimental results

Research questions

• RQ1 Can AnoSeg directly learn to generate accurate anomaly maps from synthetic anomalies and normal data without ground-truth labels?
• RQ2 Does adversarial learning with normal-region focus and coordinate channels improve segmentation accuracy beyond reconstruction-based and GradCAM-based methods?
• RQ3 To what extent does coordinate-channel information improve localization of location-dependent anomalies?
• RQ4 Can the generated anomaly maps be effectively used to boost existing anomaly detection frameworks?

Key findings

• AnoSeg produces anomaly maps that closely resemble ground-truth masks and outperforms state-of-the-art methods on the MVTec AD dataset for anomaly segmentation.
• The method achieves higher IoU and pixel-level AUROC compared to conventional SOTA approaches without relying on pretrained models.
• Ablation studies show that hard augmentation, adversarial learning with masked reconstruction, and coordinate channels each contribute to improvements in AUROC and IoU.
• The anomaly detector using AnoSeg’s maps maintains competitive image-level AUROC, demonstrating the utility of the generated maps for detection as well as segmentation.
• AnoSeg demonstrates robustness to thresholding in IoU evaluation, maintaining superior performance across a range of thresholds.