Promoting Accurate Image Reconstruction via Synthetic Noise for Unsupervised Screw Anomaly Detection and Location
Journal
IEEE Transactions on Instrumentation and Measurement
ISSN
0018-9456
1557-9662
Date Issued
2025
Author(s)
Abstract
With residuals-based unsupervised models, defects can be detected using only normal samples, boosting both efficiency and practicality in industrial anomaly detection (AD) applications. While these approaches offer benefits, they also present challenges in detecting anomalies in cylindrical objects, such as imprecise image reconstruction and camera blind spots. As a result, this paper introduces ScrewNet, an unsupervised anomaly detection framework designed specifically for cylindrical objects like screws. To improve the accuracy of reconstruction in the AD network, the synthetic noise module and pair training approach are employed during the training stage. The synthetic noise module, utilizing one of the CycleGAN generators, produces defects that match real-world distributions, thereby enabling the AD network to extract valuable features while preserving image structure. The pair training approach is designed to enable the AD network to achieve accurate reconstruction using similar background input-output pairs. To tackle inspection blind spots problem, an efficient, cost-effective inspection system using multiple mirrors is developed for the inference stage. This approach enables effective anomaly detection across the entire 360-degree surface of the screw. In experiments, three different types of screw datasets containing both damaged and undamaged screws are provided. Numerous experiments demonstrate the superiority of our approach over several state-of-the-art (SOTA) unsupervised AD networks. For instance, ScrewNet outperforms baselines by 4 dB in average Peak signal-to-noise ratio (PSNR) for image reconstruction quality and by 32% in average Area Under the Receiver Operating Characteristic curve (AUROC) for pixel-level anomaly segmentation. Additionally, image-level detection achieves 98.3% accuracy across the three screw datasets. The code is available at https://github.com/vitowen9580/ScrewNet.
Subjects
automated optical inspection system
CycleGAN-based noise synthesis
image reconstruction enhancement
Unsupervised screw anomaly detection
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Type
journal article