Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Scalable IoT Deployment for Visual Anomaly Detection via Efficient Compression

Arianna Stropeni, Francesco Borsatti, Manuel Barusco, Davide Dalle Pezze, Marco Fabris, Gian Antonio Susto

TL;DR

The paper tackles visual anomaly detection in resource-constrained IoT/IIoT settings by proposing a resource-aware pipeline that blends edge compression strategies (image and feature-based) with server-side anomaly detection. It formalizes the problem under edge budgets and demonstrates two deployment modes, supported by compression techniques such as WebP, random sampling, and product quantization. Through experiments on the MVTec AD benchmark, it shows that WebP-based feature compression can preserve high detection performance while dramatically reducing end-to-end latency (up to 80%), enabling scalable industrial deployment. The authors release MoViAD and outline future directions including adaptive bandwidth-aware strategies, online updates, and security enhancements for heterogeneous edge-server networks.

Abstract

Visual Anomaly Detection (VAD) is a key task in industrial settings, where minimizing operational costs is essential. Deploying deep learning models within Internet of Things (IoT) environments introduces specific challenges due to limited computational power and bandwidth of edge devices. This study investigates how to perform VAD effectively under such constraints by leveraging compact, efficient processing strategies. We evaluate several data compression techniques, examining the tradeoff between system latency and detection accuracy. Experiments on the MVTec AD benchmark demonstrate that significant compression can be achieved with minimal loss in anomaly detection performance compared to uncompressed data. Current results show up to 80% reduction in end-to-end inference time, including edge processing, transmission, and server computation.

Towards Scalable IoT Deployment for Visual Anomaly Detection via Efficient Compression

TL;DR

The paper tackles visual anomaly detection in resource-constrained IoT/IIoT settings by proposing a resource-aware pipeline that blends edge compression strategies (image and feature-based) with server-side anomaly detection. It formalizes the problem under edge budgets and demonstrates two deployment modes, supported by compression techniques such as WebP, random sampling, and product quantization. Through experiments on the MVTec AD benchmark, it shows that WebP-based feature compression can preserve high detection performance while dramatically reducing end-to-end latency (up to 80%), enabling scalable industrial deployment. The authors release MoViAD and outline future directions including adaptive bandwidth-aware strategies, online updates, and security enhancements for heterogeneous edge-server networks.

Abstract

Visual Anomaly Detection (VAD) is a key task in industrial settings, where minimizing operational costs is essential. Deploying deep learning models within Internet of Things (IoT) environments introduces specific challenges due to limited computational power and bandwidth of edge devices. This study investigates how to perform VAD effectively under such constraints by leveraging compact, efficient processing strategies. We evaluate several data compression techniques, examining the tradeoff between system latency and detection accuracy. Experiments on the MVTec AD benchmark demonstrate that significant compression can be achieved with minimal loss in anomaly detection performance compared to uncompressed data. Current results show up to 80% reduction in end-to-end inference time, including edge processing, transmission, and server computation.
Paper Structure (15 sections, 8 equations, 4 figures, 3 tables)

This paper contains 15 sections, 8 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and problem formulation
  3. Mathematical preliminaries
  4. Problem formulation
  5. Proposed method
  6. Deployment architecture
  7. VAD modeling and training
  8. Feature Compression Techniques
  9. Experimental results
  10. Setting and data set
  11. Examined scenarios
  12. Anomaly detection performance
  13. Inference latency
  14. Performance vs. Efficiency Trade-off
  15. Conclusions and future directions

Figures (4)

  • Figure 1: Overview of the VAD pipeline for IoT scenarios: edge devices acquire raw asset images and run compression techniques or apply feature extraction. Data are then sent across a bandwidth-limited IIoT channel to a central server. An anomaly detector (PatchCore in this study) is used to identify anomalous samples and localize defects.
  • Figure 2: Illustration of the visual anomaly detection system architectures evaluated in this study. Different compression strategies are applied at the edge to raw data and features before transmission to the server for decoding and feature extraction.
  • Figure 3: Pixel-level F1-scores across all MVTec categories.
  • Figure 4: Trade-off between pixel-level F1-score and average payload size for each method.