Technical Summary: Trustworthy Visual Quality Inspection under Data Scarcity in Manufacturing

Problem Statement

Automated visual inspection in manufacturing faces two critical barriers to economic viability and trustworthiness: data scarcity and unreliable decision-making.

1. Data Scarcity: In optimized production lines, the vast majority of units are non-defective. Consequently, labeled defect samples are rare, limiting the performance of supervised learning-based inspectors which typically require substantial training data.
2. Unreliable Decisions: Standard classifiers often output "hard labels" without confidence estimates. This creates an asymmetric cost structure: a false reject wastes a good product, while a false accept allows defective products to pass. Furthermore, models trained on limited data may confidently misclassify out-of-distribution defects, leaving manufacturers unable to distinguish between reliable predictions and uncertain ones.

Methodology

The authors propose a staged inspection pipeline that integrates synthetic data generation, unsupervised anomaly detection, and uncertainty-aware classification. The system operates in two phases: an offline training phase and an online inference phase.

1. Synthetic Defect Generation (Offline)

To address data scarcity, the authors utilize DualAnoDiff, a dual-branch diffusion model based on Stable Diffusion v1.5.

• Mechanism: Unlike traditional inpainting, DualAnoDiff jointly generates a global view of the anomalous object and a foreground view of the isolated defect region in a single denoising process. The branches share attention mechanisms to ensure consistency between the synthesized anomaly and its generated mask.
• Training: The model is fine-tuned per defect type (fingerprint, interrupted_print, smudged) using a limited set of real samples (50 images total). The best checkpoint is selected based on IC-LPIPS and KID metrics against the training pool.
• Output: For each defect category, 500 synthetic images and corresponding pixel-aligned masks are generated to augment the training dataset.

2. Unsupervised Anomaly Detection (Online)

The pipeline begins with region-of-interest (ROI) segmentation using Mask R-CNN to isolate the shaver chassis from the background.

• Model: EfficientAD is employed for anomaly detection. It is trained exclusively on defect-free samples, making it robust to the scarcity of defect data.
• Function: It uses a lightweight student-teacher architecture coupled with an autoencoder to detect deviations from normality. It outputs a scalar anomaly score and a pixel-level heatmap.
• Logic: Units scoring below a threshold are accepted as defect-free. Flagged units proceed to the classification stage. This stage is agnostic to specific defect taxonomies, focusing solely on the presence and location of anomalies.

3. Defect Classification with Uncertainty Estimation (Online)

For units flagged as anomalous, the system classifies the defect type while quantifying uncertainty. Two approaches are compared:

• Deep Ensembles: Five independently trained EfficientNet-B3 models with different random seeds. Uncertainty is derived from the disagreement (variance) among the ensemble members.
• Bayesian Neural Networks (BNN): An EfficientNet-B3 architecture modified to approximate a BNN. Early layers remain deterministic for stable feature extraction, while deeper layers utilize Flipout layers to apply pseudo-independent perturbations to weights during forward passes.
• Uncertainty Types: The system estimates both aleatoric uncertainty (observation noise) and epistemic uncertainty (model lack of knowledge). High epistemic uncertainty triggers a "defer" action, routing ambiguous units to human review rather than forcing a potentially incorrect automated decision.

Key Contributions

1. Staged Pipeline Architecture: A novel integration of ROI segmentation, unsupervised anomaly detection (EfficientAD), and uncertainty-aware classification, designed to handle the specific constraints of production lines.
2. Empirical Study on Synthetic Data: An evaluation of diffusion-based synthetic defect generation (DualAnoDiff) under extreme data scarcity, demonstrating its impact on classification and localization performance using real production defects.
3. Uncertainty-Aware Classification: A comparative analysis of Deep Ensembles and Bayesian Neural Networks for defect classification in low-data regimes, assessing predictive performance, calibration, and the ability to identify out-of-distribution samples for human intervention.

Experimental Results

The study was conducted on images of Philips electric shaver chassis, focusing on three defect categories: fingerprint, smudged, and interrupted print. Experiments were run on training sets with very few real samples (10–15 per class), supplemented by synthetic data.

• Classification Performance:
  • On a binary classification task (fingerprint vs. smudged), both Deep Ensemble and Bayesian models achieved perfect accuracy (1.000) on the small real test set (16 images).
  • On a three-class task, the Bayesian approach outperformed the Deep Ensemble as training data increased (500 synthetic images per class), achieving higher accuracy (0.833 vs. 0.750) and F1-score (0.830 vs. 0.709), as well as a superior AUROC (0.977 vs. 0.930).
• Calibration and Uncertainty:
  • The Bayesian classifier demonstrated significantly better calibration, with Brier scores and Expected Calibration Error (ECE) reduced by up to two orders of magnitude compared to the Deep Ensemble in the binary task.
  • Epistemic Uncertainty: In the binary task, the Bayesian model reported near-zero epistemic uncertainty, attributed to the high consistency of its stochastic forward passes on a small, well-represented test set. Conversely, the Deep Ensemble reported non-zero uncertainty due to inter-model disagreement.
  • Aleatoric Uncertainty: Generally decreased with more training data in the binary task, suggesting better calibration of inherent class ambiguity.
• System Latency:
  • The Deep Ensemble offered lower latency (0.0629s for classification) and higher throughput (15.90 img/s) compared to the Bayesian approach (0.5073s, 1.97 img/s) due to the computational overhead of Monte Carlo sampling in the BNN.
  • Despite the Bayesian overhead, the total pipeline latency (approx. 1.0s) remains suitable for online production-line deployment.

Significance and Claims

The authors position this work as a work-in-progress report on preliminary results. They claim that:

• Synthetic augmentation combined with uncertainty-aware classification can lower the cost of developing trustworthy, deployable inspection models in data-scarce environments.
• The proposed pipeline enables selective human intervention. By deferring only uncertain predictions to human experts, the system preserves the labor savings of automation while mitigating the risks of false accepts/rejects.
• The approach bridges the gap between the need for high accuracy and the reality of limited defect data, offering a pathway to "trustworthy" automation where the system knows when it does not know.

Future work involves evaluating the end-to-end online pipeline and implementing a feedback loop where out-of-distribution units trigger few-shot generation and retraining for new defect types.