Efficient Computer Vision for Edge AI Applies Quantization Strategies that Optimize Memory, Latency, and Accuracy while Preserving Defect Sensitivity in Industrial Anomaly Detection under Limited Power Budgets

Abstract

Limitations on industrial anomaly detection in computer vision is moving to edge devices, where memory, compute throughput, latency, and power budgets often make full-precision deep models impractical. This paper reviews how edge constraints shape deployment choices and motivates model-compression pipelines that preserve defect sensitivity while enabling real-time inference. It surveys neural-network quantization as a core strategy for reducing model size and accelerating inference, covering post-training quantization and quantization-aware training, along with uniform and mixed-precision approaches and recent methods designed to minimize accuracy loss at low bit-widths. It then summarizes industrial anomaly-detection paradigms (unsupervised and supervised), common datasets and evaluation practices, and representative methods that rely on feature backbones and efficient scoring/localization. Finally, it connects algorithmic choices to practical deployment workflows on embedded platforms, highlighting toolchains and runtimes (e.g., INT8 execution paths) that translate quantized models into measurable gains in throughput and energy efficiency. A defect-sensitive 8-bit quantization perspective is discussed to illustrate how task-aware compression can retain anomaly-detection performance while meeting strict on-device constraints.