cs.CVMar 8, 2026

Integration of deep generative Anomaly Detection algorithm in high-speed industrial line

Authors: Niccolò Ferrari, Nicola Zanarini, Michele Fraccaroli, Alice Bizzarri, Evelina Lamma

Organizations: Department of Engineering, University of Ferrara, Via Saragat 1, Ferrara, 44122, Italy. · Bonfiglioli Engineering, Via Amerigo Vespucci 20, Ferrara, 44124, Italy.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Industrial visual inspection in pharmaceutical production requires high accuracy under strict constraints on cycle time, hardware footprint, and operational cost. Manual inline inspection is still common, but it is affected by operator variability and limited throughput. Classical rule-based computer vision pipelines are often rigid and difficult to scale to highly variable production scenarios. To address these limitations, we present a semi-supervised anomaly detection framework based on a generative adversarial architecture with a residual autoencoder and a dense bottleneck, specifically designed for online deployment on a high-speed Blow-Fill-Seal (BFS) line. The model is trained only on nominal samples and detects anomalies through reconstruction residuals, providing both classification and spatial localization via heatmaps. The training set contains 2,815,200 grayscale patches. Experiments on a real industrial test kit show high detection performance while satisfying timing constraints compatible with a 500 ms acquisition slot.

Figures & tables

Figure 1: (a) Original vial-region image XX . (b) Same image with superimposed Perlin noise XnX_{n} . (c) Isolated Perlin noise map NN . (d) Binary mask MM of the superimposed region.
Figure 2: Three residual blocks in the encoder network. Only the last one halve the (H,W)(H,W) size of the layer
Figure 3: Three residual blocks in the decoder network. Only the last one double the (H,W)(H,W) size of the layer
Figure 4: The product: a BFS strip composed by 5 vials 1 .
Figure 5: The vial regions 1 : the logic regions in which the vial is divided.
Figure 6: Three residual blocks in the encoder (ResEnc). One stage consists of three convolutional blocks in the order A, B, C.
Figure 7: The bottleneck stage.
Figure 8: Three residual blocks in the decoder (ResDec). One stage consists of three transposed-convolution blocks in the order A, B, C.
Figure 9: The training flowchart.
Overall results
 R0R1R2R3Vial
Threshold0.0155890.0380170.0465680.029593-
Accuracy0.99190.99260.99570.99910.9870
True positive ratio0.99660.99850.99860.99940.9942
True negative ratio0.90930.90440.97790.99730.9581
Balanced accuracy0.95300.95150.98830.99840.9762
Table 1: Overall results using the equation 10 as anomaly score.
Per strip results
 Score
Threshold R00.016156
Threshold R10.038584
Threshold R20.046635
Threshold R30.029660
Accuracy0.9593
Table 2: Per strip results using the equation 14 as anomaly score.
Per strip results
 Score
Threshold R00.013222
Threshold R10.034650
Threshold R20.046201
Threshold R30.029226
Accuracy0.9641
Table 3: Per run results using the equation 14 as anomaly score.
Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 10: Stuck particle in the upper part of the product
Figure 11: Stuck particle and anomalous liquid behavior that results in foam formation because of contamination
Figure 12: Stuck particle and anomalous liquid behavior that results in foam formation because of contamination
Figure 13: Lower body deformation
Figure 14: Light lower body deformation
Figure 15: Lower body deformation
Figure 16: Black spot on upper part of the vial’s body.
Figure 17: Burn in the lower part of the vial.
Figure 18: Light scratch near the product neck.
Figure 19: External imperfection caused by laser cutting of the product.
Figure 20: Deep scratch on the upper part of the vial’s body.
Figure 21: Burn on the vial’s body. It is reproduced as a bubble, since the network does not know how to represent the defect features.
Figure 22: Black spot on the vial’s body
Figure 23: Black spot of the vial’s body
Figure 24: Bump on the lower part of the vial’s neck, near to a big bubble stuck in the internal layer of the BFS surface.
Figure 25: Light bump on the vial’s neck
Figure 26: Heavy deformation near the neck region, overlapped to a bubble that the network is still able to reproduce.
Figure 27: Scratch on the vial’s neck.
Figure 28: Light bump on the lower part of the vial’s cap region.
Figure 29: Scratch on the vial’s neck.

Explore similar work

CardsList

  1. GRD-Net: Generative-Reconstructive-Discriminative Anomaly Detection with Region of Interest Attention Module

    Mar 8, 2026Niccolò Ferrari, Michele Fraccaroli, Evelina LammaIndustrial Anomaly DetectionAnomaly Localization

    Department of Engineering, University of Ferrara, Via Saragat 1, 44122 Ferrara, Italy · Bonfiglioli Engineering, Via Amerigo Vespucci 20, 44124 Ferrara, Italy

  2. UniScale: Arbitrary-Scale Industrial Anomaly Generation

    Aug 8, 2026Shilei Zeng, Linxin Guan, Xurui Li +2Industrial Anomaly DetectionSynthetic Anomaly Generation

    School of Electronic Information and Communications, Huazhong University of Science and Technology

  3. DeCo: Zero-Shot Industrial Anomaly Generation through Decoupling and Recoupling

    Aug 8, 2026Shilei Zeng, Xurui Li, Yaohan Tang +1Industrial Anomaly DetectionSynthetic Anomaly Generation

    School of Electronic Information and Communications, Huazhong University of Science and Technology