Efficiency-Optimized PatchCore for Anomaly Detection in Future Edge Deployment

Anomaly detection plays a critical role in industrial quality control, ensuring defect-free production in sectors like automotive, electronics, and manufacturing. Traditional anomaly detection methods rely on supervised learning, which requires large datasets with labeled anomalies—often impractical in real-world scenarios. Unsupervised methods such as PatchCore have gained attention for their ability to detect anomalies without labeled defect data, making them particularly valuable for automated inspection systems in factories. Despite its high accuracy, PatchCore has limitations that hinder its deployment in real-time edge devices like smart cameras used in industrial environments. The memory bank storing patch-level embeddings grows exponentially with dataset size, making it computationally expensive and requiring large storage capacity. This restricts its practical application in low-memory, low-power industrial devices. To address these limitations, this thesis explores a memory-efficient optimization of PatchCore by integrating Principal Component Analysis (PCA) and K-Means clustering. PCA is applied to reduce feature dimensionality, preserving the most informative components while eliminating redundancy. K-Means clustering further compresses the memory bank by grouping similar feature vectors and using cluster centroids instead of all stored patches. This approach significantly reduces memory usage while maintaining a balance between accuracy and computational efficiency. The proposed PCA-KMeans PatchCore method was evaluated on the MVTec Anomaly Detection Dataset, a benchmark for industrial defect detection. Results show that, compared to the original PatchCore, the proposed method achieves a substantial reduction in memory footprint while maintaining competitive anomaly detection performance. Although a slight drop in accuracy was observed, the efficiency gains make the model more practical for real-time applications in smart manufacturing. This study demonstrates that optimizing PatchCore for future edge deployment is feasible, making real-time anomaly detection on resource-constrained devices a practical reality.