WEDGE-Net: Wavelet-Driven Memory-Efficient Anomaly Detection for Industrial Edge Computing

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning-based anomaly detection (AD) faces challenges in industrial edge deployment due to memory and computational constraints.
• Existing methods often prioritize accuracy over operational efficiency, neglecting real-world factors like noise and latency.

Purpose of the Study:

• To develop an efficient and accurate anomaly detection model for resource-constrained edge devices in industrial settings.
• To address the limitations of current AD approaches by balancing structural precision with extreme memory efficiency.

Main Methods:

• Introduced WEDGE-Net, a dual-stream architecture decoupling anomaly detection into Frequency (DWT) and Context streams.
• The Frequency Stream filters environmental noise, while the Context Stream uses a Semantic Module for feature extraction and object consistency.
• Synthesized both streams to suppress noise and enhance structural feature compactness.

Main Results:

• WEDGE-Net achieved 97.82% mean image-level AUROC on the MVTec AD dataset with 1% memory compression.
• Demonstrated superior noise resistance in robustness analysis on the 'Tile' category.
• Achieved 686.5 FPS inference speed on an RTX 4090 GPU, with a 2.1x speedup over PatchCore-10% while maintaining competitive accuracy.

Conclusions:

• WEDGE-Net offers a practical solution for real-time industrial inspection on edge devices.
• The model effectively balances high detection accuracy with extreme memory efficiency and operational speed.
• This work provides a valuable reference for deploying advanced AD systems in manufacturing environments.