A benchmark for IoT-deployed visual anomaly detection under compression constraints

Visual Anomaly Detection (VAD) is a critical task in industrial quality control, where early identification of defective parts directly impacts operational efficiency. While deep learning has enabled highly accurate VAD systems, deploying such models within Internet of Things (IoT) environments remains challenging due to the limited computational capacity and communication bandwidth of edge devices. This study investigates the deployment of VAD models in bandwidth-constrained Industrial IoT (IIoT) environments across four VAD algorithms: PatchCore, PaDiM, FastFlow, and Dinomaly. In particular, this study provides a systematic analysis of the relationships among detection performance, computational footprint, communication overhead, and end-to-end transmission latency, highlighting the key trade-offs that emerge when deploying VAD pipelines in resource-constrained IoT environments. The framework decouples the detection pipeline into two stages: an edge-side compression or feature extraction step, and a server-side anomaly detection step. We systematically evaluate two families of transmission strategies, image compression and feature-level compression. Experiments on the MVTec AD benchmark demonstrate that WebP image compression achieves the best accuracy–efficiency trade-off, retaining detection performance close to the uncompressed baseline while reducing transmitted payload from 61.76 kB to 1.40 kB. Overall, our results show that end-to-end inference latency can be reduced by up to 80% compared to uncompressed pipelines, while maintaining competitive performance, demonstrating the viability of scalable VAD deployment in real-world IIoT scenarios.