Search for a command to run...
As deep learning-based Anomaly Detection (AD) transitions from theoretical research to industrial application, the focus is shifting towards operational efficiency and economic viability on edge devices. While recent studies have achieved remarkable detection accuracy on standard benchmarks, they often rely on heavy memory banks or complex backbones, which pose challenges for deployment in resource-constrained manufacturing environments. Furthermore, real-world inspection lines often present distinct challenges—such as environmental noise and strict latency requirements—that are not fully addressed by accuracy-centric metrics. To bridge the gap between high-performance research models and practical edge deployment, we introduce WEDGE-Net. Our approach is designed to balance structural precision with extreme memory efficiency. We decouple anomaly detection into two specialized streams: (1) a Frequency Stream (DWT) that physically filters out environmental noise to isolate structural defects, and (2) a Context Stream where a Semantic Module explicitly guides feature extraction to enforce object consistency. By synthesizing these two modalities, WEDGE-Net effectively suppresses high-frequency noise while enhancing structural-feature compactness. To validate operational stability, we conducted a robustness analysis of the ‘Tile’ category, which poses a challenging task for distinguishing defects from high-frequency textures. In this stress test, WEDGE-Net demonstrated superior resistance to environmental noise compared to conventional methods. Experimental results on the MVTec AD dataset demonstrate that WEDGE-Net achieves a mean image-level AUROC of 97.82% and an inference speed of 686.5 FPS (measured on an RTX 4090 GPU) under an extreme 1% memory-compression setting. Notably, our method demonstrates superior efficiency, achieving a 2.1× inference speedup over the widely adopted comparative model (PatchCore-10%) while maintaining competitive detection accuracy (e.g., 100% AUROC on Transistor). We hope this work serves as a practical reference for implementing real-time industrial inspection on resource-constrained edge devices.