Leveraging Unsupervised Learning for Cost-Effective Visual Anomaly Detection

Technical Summary: Leveraging Unsupervised Learning for Cost-Effective Visual Anomaly Detection

Problem Statement
Traditional machine learning-based visual inspection systems for Small and Medium-sized Enterprises (SMEs) face significant barriers to adoption. These systems typically require extensive data collection, repetitive model training, expensive hardware (high-end GPUs and cameras), and specialized machine learning expertise. Furthermore, deep learning approaches often rely on large-scale, labeled datasets of both normal and defective products. In industrial settings, the diversity of defects and the need for continual data annotation make these requirements burdensome for SMEs, which often lack the capacity to collect and maintain sufficient image data. Additionally, standard deep learning models struggle to generalize to new products or defect types without retraining, further limiting their practical utility in flexible manufacturing environments.

Methodology
This study proposes a cost-effective visual anomaly detection system designed for resource-constrained environments, specifically utilizing a Raspberry Pi 4B (8GB RAM) and a low-cost camera. The core methodology leverages unsupervised learning, where models are trained exclusively on images of "normal" products, eliminating the need for labeled defect data.

• Algorithms: The research evaluates four unsupervised learning algorithms from the Anomalib library: PaDiM (Patch-wise Anomaly Detection via Density Modeling), PatchCore, CFlow-AD, and FastFlow.
• Deployment: The models are optimized for edge deployment using the OpenVINO toolkit, which reduces model size and computational costs while maintaining accuracy.
• Experimental Setup: The system was tested on gearbox components. The workflow involved capturing images of normal parts to train the models.
  • Training Data: Initially, 20 images of normal products were used, with 15 randomly selected for training and 5 for testing.
  • Testing Conditions: The system was evaluated under various "product condition" anomalies (e.g., missing gear teeth, missing parts, extra gears, un-anodized casings) and "setup condition" anomalies (e.g., lighting changes, tray misalignment, part tilting).
• Metrics: Performance was measured using AUROC (Area Under the Receiver Operating Characteristic Curve) and F1 Macro scores. An ablation study was conducted to compare performance across different dataset sizes (20, 50, and 80 images) and hardware platforms (Raspberry Pi vs. Intel i7 PC).

Key Contributions

1. Low-Cost System Development: The creation of a functional visual anomaly detection system using a Raspberry Pi 4B, demonstrating an affordable method for detecting defects in gearbox parts without high-end computing infrastructure.
2. Algorithm Evaluation: A comparative analysis of multiple pre-trained unsupervised learning algorithms (PaDiM, PatchCore, CFlow-AD, FastFlow) specifically for deployment on low-cost hardware.
3. Performance Validation: Empirical validation of the system's ability to detect various defect types and setup variations in a laboratory setting, including the generation of heatmaps for defect localization and confidence scores for decision support.

Results

• High Accuracy with Minimal Data: On the Raspberry Pi, the system achieved an F1 macro score exceeding 0.95 and an AUROC of 1.0 across all tested product and setup conditions when trained on only 10 normal images (using 15 for training in the initial setup). The system successfully distinguished between normal and anomalous parts without prior labeling of defects.
• Hardware Constraints:
  • PaDiM proved to be the most efficient model, training in approximately 1 minute and 9 seconds on the Raspberry Pi with 20 images, achieving an F1 score of 0.95.
  • PatchCore also performed well on small datasets (20 images) but encountered system crashes when the training set size increased to 50 or 80 images due to memory limitations.
  • CFlow-AD and FastFlow failed to train successfully on the Raspberry Pi across all dataset sizes, resulting in system crashes.
• Inference Speed: While training was rapid for PaDiM, inference times were relatively high for the target application: approximately 24 seconds for PaDiM and 70 seconds for PatchCore on the Raspberry Pi.
• Sensitivity to Environment: The system showed sensitivity to environmental changes. For instance, under extremely dark conditions, the model misclassified a normal part as anomalous with only 50% confidence. Additionally, the system struggled to precisely localize subtle defects (e.g., slightly blurred missing gear teeth) and occasionally confused setup changes (like lighting or tray movement) with product defects.

Significance and Claims
The paper claims to present a pioneering exploration of deploying state-of-the-art unsupervised deep learning models on low-cost, resource-constrained hardware. It demonstrates that SMEs can achieve high-accuracy anomaly detection (F1 > 0.95) using minimal data (as few as 10 normal images) and affordable hardware, significantly lowering the barrier to entry for automated visual inspection.

However, the authors maintain a modest stance regarding the system's current limitations. They acknowledge that the system is not yet suitable for high-frequency production lines due to inference times (24–70 seconds) and its sensitivity to environmental noise (lighting, positioning). The study concludes that while the approach is swift and economical, future work is required to improve reliability through model distillation or data compression to better handle the large volume of data and limited resolution inherent in low-cost edge devices. The system is positioned as a viable solution for SMEs where rapid, high-throughput inspection is not the primary constraint, or as a supplementary check rather than a standalone replacement for human inspection in complex, variable environments.