Source-linked AI summary
Anomaly Detection on Small Industrial Components via Vision-Based Tactile Sensing
G. F. Preziosa, M. Casiglia, M. Faroni, A. M. Zanchettin, P. Rocco
TL;DR
Small industrial components can contain geometry-driven defects that RGB inspection may resolve poorly, while supervised defect data are difficult to collect. This paper benchmarks four unsupervised feature-embedding methods on GelSight tactile data from five components across nominal-sample, position-transfer, and resolution conditions. The results support tactile sensing as a low-cost, flexible quality-control alternative, while contact wear remains a deployment constraint.
Problem
Industrial anomaly detection lacks established evidence for tactile inspection of small real-world components whose geometric defects may be weakly resolved in RGB images.
Method
The study benchmarks SPADE, PaDiM, FAPM, and InReaCh on five GelSight Mini tactile datasets under good-fraction, cross-position, and low- versus high-resolution validations.
Results
Across deployment conditions, SPADE provides the strongest localization and zero-shot cross-position robustness, although it requires the full nominal pool.
Takeaways & Limitations
Vision-based tactile sensing combined with unsupervised feature embeddings can serve as a low-cost and flexible alternative to dedicated quality-control machines for small components.
Takeaways & Limitations
Contact-based sensing is constrained by soft-gel wear, contact-cycle throughput, and the need to select suitable small sensing regions.
Abstract
from arXiv · showhide
Automated inspection of small industrial components, including sub-centimetre-scale parts where defects are geometry-driven and poorly resolved by standard optical cameras, calls for sensing modalities that can directly capture fine surface geometry. Vision-based tactile sensors address this need by converting contact imprints into high-resolution image-like data compatible with existing deep-learning pipelines, yet their effective use for industrial anomaly detection (AD) remains largely unexplored. This work systematically evaluates unsupervised AD methods on a real tactile dataset covering five genuine industrial components acquired with a GelSight Mini sensor mounted on a collaborative robot. Four feature-embedding methods, SPADE, PaDiM, FAPM, and InReaCh, are compared under three validations explicitly motivated by the deployment constraints of contact-based sensing: a Good Fraction analysis establishing the minimum number of nominal contacts for stable performance, directly bounded by gel wear since every acquisition degrades the soft interface; a cross-position evaluation assessing generalization across different contact locations observing the same recurring surface pattern; and a low- versus high-resolution comparison evaluating the cost-benefit of higher-resolution tactile acquisition. Overall, this systematic benchmarking study provides practical guidance for researchers and practitioners adopting vision-based tactile sensing for industrial AD and shows how this modality can serve as a viable alternative for industrial quality-control tasks.
1. Introduction
Industrial quality control needs flexible, low-downtime detection of subtle defects, but nominal-only anomaly detection is more practical than supervised learning when defects are rare and heterogeneous. This work evaluates tactile sensing as an image-compatible way to capture geometric defects that RGB cameras may miss.
- Motivation: Industrial inspection must detect subtle defects reliably while limiting downtime, manual intervention, false alarms, and missed-defect consequences.False positives can interrupt production, while missed defects may propagate to assembly, warranty, or safety-critical stages.
- Industrial anomaly detection: Unsupervised anomaly detection models nominal production data and identifies deviations without assuming fixed defect classes.Its outputs can support item-level decisions or spatial defect localization.
- Sensing gap: RGB-based inspection may poorly resolve geometry-driven defects on very small components, including burrs, dents, flattening, thread irregularities, and edge damage.Manual inspection is time-consuming and subjective, while dedicated machines can be expensive and inflexible.
- Study scope: The study systematically evaluates vision-based tactile sensing for industrial anomaly detection using unsupervised methods under tactile-specific acquisition constraints.The comparison varies nominal-image count, cross-position generalization, and low- versus high-resolution acquisition.
2. Related Work
Industrial anomaly detection has been developed mainly for optical RGB datasets and methods, whereas tactile sensing offers contact-derived geometric information but introduces wear, throughput, and sensing-area constraints. The study therefore focuses on feature-embedding methods as a practical nominal-only approach for tactile inspection.
- Vision-based tactile sensing: Vision-based tactile sensors produce RGB images of local contacted surfaces, but repeated contact wears soft gels and limits throughput and sensing-area flexibility.These constraints motivate evaluating required nominal contacts and tactile image resolution.
- Benchmarks: Established unsupervised industrial anomaly-detection benchmarks use optical RGB images and primarily evaluate appearance or texture deviations.No established tactile dataset covers unsupervised industrial anomaly detection on small real-world components.
- Anomaly-detection methods: Feature-embedding methods represent nominal images or patches in latent space using boundaries, distributions, or stored exemplars.The family includes one-class, Teacher–Student, distribution-modeling, and memory-bank methods.
- Method rationale: The study selects feature-embedding methods because they typically require fewer nominal training images than reconstruction or synthesized-anomaly approaches.Fewer images reduce contacts and associated gel wear while supporting adaptation to new component geometries.
3. Tactile Inspection Dataset
The tactile dataset contains five heterogeneous industrial components with varied geometries, defect types, resolutions, and acquisition positions. Samples were acquired with a robot-mounted GelSight Mini using both constrained and manually positioned fixtures, alongside artificially generated and some genuine defects.
- Industrial components: The dataset covers five real components, PZ1–PZ5, selected for difficult inspection conditions and heterogeneity in size and defect typology.The components range from threaded or grooved parts to sub-centimetre-scale components.
- Defect generation: Defects are artificially induced with tools such as pliers, hammers, and sandpaper under domain-expert supervision to reproduce plausible industrial damage.PZ2 also contains genuine industrial defects.
- Acquisition protocol: A GelSight Mini mounted on a FANUC CRX-10iA collaborative robot acquired the tactile data using fixtures for PZ1–PZ3 and manual placement for PZ4–PZ5.Two gels were consumed over the complete acquisition campaign.
- Dataset composition: All components were acquired at 320×240 resolution, while PZ3–PZ5 also used 1280 × 720; acquisition positions ranged from one to four depending on component geometry.PZ1 used four poses, PZ3 two, and PZ2, PZ4, and PZ5 one pose.
4. Implemented Feature-Embedding Methods
The study compares four feature-embedding anomaly-detection methods that learn nominality from frozen pretrained deep features. They differ in whether they use nearest-neighbor descriptors, spatial Gaussian models, prototype memories, or related feature-space representations, with tiled processing for high-resolution tactile images.
- Compared methods: SPADE, PaDiM, FAPM, and InReaCh are compared as feature-embedding methods using pretrained backbone features from nominal samples.Their shared principle is modeling normality in feature space.
- SPADE: SPADE compares global and multi-scale local descriptors with nearest nominal images and patches for image-level scoring and localization.Image-level comparison uses the 5 nearest nominal images.
- PaDiM: PaDiM models each spatial location with a multivariate Gaussian over concatenated backbone features and scores deviations using regularized Mahalanobis distances.The concatenated descriptor has 1792 dimensions, with 550 channels retained for scoring.
- FAPM and InReaCh: FAPM and InReaCh use feature-space memory or prototype representations to characterize nominal tactile samples.FAPM retains 10% of local descriptors in its near prototype bank.
- Image processing: All methods use a frozen ImageNet-pretrained Wide-ResNet-50-2, while high-resolution images are processed as overlapping 224 × 224 tiles and recomposed into heatmaps.Low-resolution images use 224×224 central crops; high-resolution tiles overlap by 56 pixels.
5. Experimental Comparisons
The experiments evaluate unsupervised tactile anomaly detection under nominal-data, cross-position, and resolution constraints. SPADE provides the strongest localization overall, while small target-position samples and high-resolution acquisition can improve other methods.
- Validation 1: Good Fraction Analysis: Validation 1 varies the Good Fraction from 5% to 100% across 10 training seeds to identify when nominal-data performance stabilizes.Each component uses a reference position, held-out nominal validation images, and defective samples from the same position.
- Validation 1: Good Fraction Analysis: SPADE consistently attains the highest AUC-PRO@0.01 across all five components, with the largest gap on sub-centimetre-scale PZ4 and PZ5.PaDiM forms a stable second tier on most components.
- Validation 2: Cross-Position Evaluation: Validation 2 trains at one acquisition position and tests transfer to another, varying target-position nominal injection from 0% to 15%.The evaluation covers PZ1 and PZ3, where recurrent geometry makes position transfer meaningful.
- Validation 2: Cross-Position Evaluation: 5% target-position injection clearly improves most method–component pairs, while gains from 5% to 15% are typically smaller.The main exception is AUC-PRO@0.01 on PZ3, where SPADE and InReaCh perform best at 0%.
- Validation 3: Low- versus High-Resolution Comparison: High-resolution tiled acquisition generally improves pixel-level localization, especially AUPRC and AUC-PRO@0.01, and changes the method ranking.PaDiM and FAPM improve consistently across PZ3–PZ5, while FAPM becomes most competitive overall in high resolution.
6. Conclusion
The study systematically evaluates unsupervised feature-embedding methods for anomaly detection on small industrial components sensed through vision-based tactile imaging. Its deployment comparison finds complementary method strengths rather than a single dominant method across all conditions.
- 6. Conclusion: The work systematically evaluates unsupervised feature-embedding methods on small real-world industrial components acquired with a vision-based tactile sensor.Table 4 summarizes behavior across the deployment conditions, and Fig. 5 provides representative localization outcomes.
- 6. Conclusion: SPADE offers the best localization and strongest zero-shot cross-position robustness but requires the full nominal pool.This makes SPADE less suitable when gel wear must be minimized.
- 6. Conclusion: InReaCh converges with the fewest contacts and handles position shifts without target samples, but has lower absolute accuracy.Its advantage is therefore concentrated in contact efficiency and zero-shot transfer rather than overall localization performance.
- 6. Conclusion: PaDiM and FAPM benefit most from few-shot cross-position injection and high-resolution acquisition.They are preferable when high-resolution imaging is available and a small number of target-position samples can be collected.