Global-Local Consistency for Cost-Effective Anomaly Localization with Point Annotations

• Published in: 2024 16th IIAI International Congress on Advanced Applied Informatics (IIAI-AAI) pp. 53 - 57
• Main Authors: Nakanishi, Keiichi, Katafuchi, Ryoya, Tokunaga, Terumasa
• Format: Conference Proceeding
• Language: English
• Published: IEEE 06.07.2024
• Subjects: Annotations; Anomaly detection; Costs; Deep Learning; Faces; Image Anomaly Detection; Image segmentation; Informatics; Location awareness; Medical diagnosis; Point Annotation; Semi-Supervised Learning; Semisupervised learning; Supervised learning; Visual inspection AI
• DOI: 10.1109/IIAI-AAI63651.2024.00019

Creating pixel-level ground-truth (GT) masks is quite costly for deep learning-based image segmentation. Specialists in areas such as anomaly detection and medical diagnostics face difficulties in producing many GT masks due to limited resources. To reduce this burden, we propose a cost-effective image segmentation framework with point annotations (CoSPA) that performs image segmentation with only point annotations. Point annotations refer to the partial and sparse labeled coordinates in an image. The key idea is to ensure consistency between the predictions for the same coordinates from the two different networks of CoSPA. This new consistency enables the CoSPA to improve segmentation performance and extend to semi-supervised learning. For the MVTec AD dataset, we verified the cost-effectiveness of CoSPA through an anomaly detection task. We demonstrated that the point annotating costs were reduced by 80% compared to creating GT masks. Subsequently, the CoSPA realized by 87% of the mean Intersection over Union (mIoU) achieved using the fully supervised method, DeepLabV3+. Moreover, the mIoU of CoSPA using only 30% of all point annotations defeated that of the unsupervised method, PaDiM. This study offers a new direction for economic anomaly localization.