T-NACD: A Tactile-Friendly Novel Anomaly Class Discovery Framework for Black Rubber Products
Long Xiao,, Kailin Lyu,, Jianing Zeng, Xuexin Liu,, Zhuojun Zou, Lin Shu, Jie Hao,∗
AI summary
Problem
Visual anomaly detection methods fail on black rubber products due to low reflectance, matte textures, and undetectable hardness variations. There is no dedicated framework to discover novel anomaly classes using tactile sensing for this challenging material.
Approach
The framework adapts pre-trained visual models to tactile data using tactile-specific augmentations and a lightweight feature adapter, while employing dual-margin entropy regularization to separate overlapping anomaly classes and preserve known category knowledge.
Key results
- Surpasses visual SOTA methods by 6.9% on the new BRD tactile dataset
- Enables effective visual-to-tactile domain transfer with minimal training
- Reduces feature stocking and catastrophic forgetting via dual-margin regularization
- Achieves state-of-the-art performance on FabricVST and YCB tactile datasets
Why it matters
Provides a reliable, vision-independent inspection solution for manufacturing processes where tactile sensing is critical for quality control of dark, matte materials.
Abstract
Novel Anomaly Class Discovery (NACD) has re- cently gained attention in industrial anomaly detection, aiming not only to recognize defects but also to recognize fine- grained and previously unseen anomaly types. However, ex- isting methods, primarily based on visual images, struggle to handle common but challenging materials such as black rubber products. Such products often exhibit hardness anomalies and visually indistinguishable surface defects during production due to process instability and high material absorbance. To address this, we present T-NACD, the first tactile-friendly framework for NACD, leveraging tactile sensing to detect subtle geometric anomalies that are invisible to vision. We introduce two tactile- specific data augmentation methods and design a lightweight feature adapter to transfer visual pre-trained models to the tac- tile domain without the need for large-scale training. Moreover, we propose a dual-margin enhanced entropy regularization method to mitigate forgetting of known categories and reduce feature stocking across similar anomaly classes. To support this study, we collect BRD, the first real-world tactile dataset for black rubber anomaly detection, including BRD-A (surface and hardness defects) and BRD-H (diverse hardness anomalies). Experimental results show that T-NACD outperforms visual SOTA methods by 6.9% and generalizes well to other non- industrial tactile detection tasks.