pith. sign in

arxiv: 2409.15980 · v2 · pith:NFYPRTRRnew · submitted 2024-09-24 · 💻 cs.CV · cs.AI

Leveraging Unsupervised Learning for Cost-Effective Visual Anomaly Detection

Yunbo Long , Zhengyang Ling , Sam Brook , Duncan McFarlane , Alexandra Brintrup This is my paper

Pith reviewed 2026-05-23 20:56 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords visual anomaly detectionunsupervised learningRaspberry Picost-effective inspectionindustrial automationAnomalibopenVINOminimal training data
0
0 comments X

The pith

Unsupervised learning enables visual anomaly detection on Raspberry Pi hardware using only ten normal images.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to demonstrate that visual anomaly detection for product inspection can be performed with minimal data and low-cost equipment by applying unsupervised models. It shows a complete training and inference pipeline that runs on a Raspberry Pi in 90 seconds from just 10 normal images while reaching high accuracy. This matters because conventional systems demand large labeled datasets, costly hardware, and machine-learning specialists, barriers that exclude many small manufacturers. A reader would see the work as a route to practical factory automation without those overheads.

Core claim

The research develops a low-cost visual anomaly detection solution that uses minimal data for model training while maintaining generalizability and scalability. The system utilises unsupervised learning models from Anomalib and is deployed on affordable Raspberry Pi hardware through openVINO. The results show that this cost-effective system can complete anomaly detection training and inference on a Raspberry Pi in just 90 seconds using only 10 normal product images, achieving an F1 macro score exceeding 0.95. While the system is slightly sensitive to environmental changes like lighting, product positioning, or background, it remains a swift and economical method for factory automation.

What carries the argument

Unsupervised learning models from Anomalib, adapted and deployed via openVINO on Raspberry Pi hardware, which perform training and inference from only 10 normal images.

If this is right

  • Small and medium-sized manufacturers gain access to automated inspection without large data collections or expert teams.
  • The approach supports rapid setup for new production lines using only a handful of good samples.
  • Deployment on affordable single-board computers reduces capital costs compared with traditional inspection rigs.
  • The pipeline preserves scalability while cutting the usual requirements for repetitive model retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same minimal-data pattern could be tested on other low-power edge devices to widen hardware options.
  • Extending the method to multi-product lines would require checking whether 10 images per product remain sufficient.
  • The acknowledged sensitivity to environment suggests a natural next step of adding lightweight adaptation steps at inference time.

Load-bearing premise

The unsupervised models from Anomalib continue to deliver strong detection performance when run on Raspberry Pi hardware via openVINO even when supplied with only 10 normal images.

What would settle it

A direct test on the Raspberry Pi deployment that records substantially lower detection reliability when the same 10 training images are used but lighting, positioning, or background is altered.

Figures

Figures reproduced from arXiv: 2409.15980 by Alexandra Brintrup, Duncan McFarlane, Sam Brook, Yunbo Long, Zhengyang Ling.

Figure 1
Figure 1. Figure 1: System Workflow 3.2. Hardware The Raspberry Pi 4B was used in all tests. (Mounir et al., 2020) compared the execution time performance of basic im￾age processing algorithms on various platforms. It highlighted that the Raspberry Pi can run lightweight deep learning models efficiently and is comparable to entry-level x86 personal com￾puters (PCs). Raspberry Pis has been successfully employed in various low-… view at source ↗
Figure 2
Figure 2. Figure 2: Images of different setup conditions (b, c and d) and different product conditions (e, f, g, and h). The normal photos were then divided into training and testing datasets in a 2:1 ratio. Then, the model was trained exclusively on normal photos, while testing utilized both normal and all ab￾normal photos. This study aims to examine how varying dataset sizes (20, 40, and 80 images, with 25% being normal) im… view at source ↗
Figure 6
Figure 6. Figure 6: Aonomaly Detection Results for Gears with Missing Teeth [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 3
Figure 3. Figure 3: Aonomaly Detection Results for Non-Anodized Tilt Parts [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Aonomaly Detection Results for Normal Parts [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Misclassification in Anomaly Detection for Normal Parts [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
read the original abstract

Traditional machine learning-based visual inspection systems require extensive data collection and repetitive model training to improve accuracy. These systems typically require expensive camera, computing equipment and significant machine learning expertise, which can substantially burden small and medium-sized enterprises. This study explores leveraging unsupervised learning methods with pre-trained models and low-cost hardware to create a cost-effective visual anomaly detection system. The research aims to develop a low-cost visual anomaly detection solution that uses minimal data for model training while maintaining generalizability and scalability. The system utilises unsupervised learning models from Anomalib and is deployed on affordable Raspberry Pi hardware through openVINO. The results show that this cost-effective system can complete anomaly defection training and inference on a Raspberry Pi in just 90 seconds using only 10 normal product images, achieving an F1 macro score exceeding 0.95. While the system is slightly sensitive to environmental changes like lighting, product positioning, or background, it remains a swift and economical method for factory automation inspection for small and medium-sized manufacturers. The code is available at https://github.com/Yunbo-max/Cost-Effective-Visual-Anomaly-Detection-using-Unsupervised-Learning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript describes a cost-effective visual anomaly detection system that applies unsupervised models from the Anomalib library, optimized and deployed on Raspberry Pi hardware using OpenVINO. It reports that training and inference can be completed in 90 seconds using only 10 normal product images while attaining an F1 macro score above 0.95, positioning the approach as accessible for small and medium-sized manufacturers despite noted sensitivity to environmental factors such as lighting and positioning. The code is made available via GitHub.

Significance. If the headline performance numbers prove robust, the work demonstrates a practical route to lowering hardware and data barriers for industrial visual inspection by repurposing existing open-source tools (Anomalib + OpenVINO) on commodity edge devices. Public code release aids reproducibility and potential follow-on use by practitioners.

major comments (2)
  1. [Abstract] Abstract: the central empirical claims (90 s training+inference on Raspberry Pi, F1 macro >0.95 with 10 normal images) are stated without any description of the test dataset (number of anomalous samples, source, or split), the specific Anomalib models selected, or statistical validation (multiple runs, error bars). These omissions are load-bearing for assessing whether the unsupervised approach actually delivers the advertised performance.
  2. [Abstract] Abstract: the text acknowledges sensitivity to lighting, product positioning, and background yet supplies no quantitative ablation, perturbation tests, or robustness metrics under controlled variations of these factors. Because the target use case is factory-floor inspection where such variability is routine and only 10 normal samples are used, the absence of this analysis directly weakens support for the generalizability claim.
minor comments (2)
  1. [Abstract] Abstract contains a clear typo: 'anomaly defection training' should be 'anomaly detection training'.
  2. The manuscript would benefit from a table or figure summarizing hardware specifications, model variants tested, and exact timing breakdowns (training vs. inference) to make the 90-second figure reproducible.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback on the abstract and the need for stronger support of the performance claims. We address each major comment below and indicate where revisions will be made.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central empirical claims (90 s training+inference on Raspberry Pi, F1 macro >0.95 with 10 normal images) are stated without any description of the test dataset (number of anomalous samples, source, or split), the specific Anomalib models selected, or statistical validation (multiple runs, error bars). These omissions are load-bearing for assessing whether the unsupervised approach actually delivers the advertised performance.

    Authors: We agree the abstract should be more self-contained. The full manuscript details the custom product-image dataset (including the number of anomalous samples and the train/test split), the specific pre-trained Anomalib models employed, and the single-run experimental protocol. In revision we will expand the abstract to briefly state the dataset source and size, name the models, and clarify that the reported F1 macro >0.95 is the primary observed value under the described conditions. revision: yes

  2. Referee: [Abstract] Abstract: the text acknowledges sensitivity to lighting, product positioning, and background yet supplies no quantitative ablation, perturbation tests, or robustness metrics under controlled variations of these factors. Because the target use case is factory-floor inspection where such variability is routine and only 10 normal samples are used, the absence of this analysis directly weakens support for the generalizability claim.

    Authors: We acknowledge that the absence of quantitative robustness metrics limits the strength of the generalizability claim. Our experiments focused on demonstrating feasibility with 10 normal images and 90-second training on Raspberry Pi; the sensitivity statement is based on informal observations rather than controlled ablations. We will revise the abstract and discussion to more explicitly frame this as a limitation and identify controlled robustness testing as future work. revision: partial

standing simulated objections not resolved
  • Quantitative ablation or perturbation results for lighting, positioning, and background variations, because no such controlled experiments were conducted.

Circularity Check

0 steps flagged

No circularity; performance claims are direct empirical measurements from standard models

full rationale

The paper reports measured runtime (90 s) and F1 macro (>0.95) obtained by running pre-existing Anomalib unsupervised models on Raspberry Pi hardware via openVINO with a fixed set of 10 normal images. No derivation chain, equations, fitted parameters, or self-referential predictions appear in the provided text. The results are presented as experimental outcomes rather than quantities derived from the setup itself. Acknowledged sensitivity to lighting/positioning is a robustness concern, not a circularity issue. The central claims therefore remain independent of any self-definition or self-citation reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The performance claim rests on the generalization ability of pre-trained unsupervised models from Anomalib when used with minimal normal samples on constrained hardware; no new entities or fitted parameters are introduced in the abstract.

axioms (1)
  • domain assumption Unsupervised anomaly detection models can learn normal patterns sufficiently from a small number (10) of normal samples to achieve high detection accuracy
    Invoked by the claim of training with only 10 normal images and achieving F1 > 0.95

pith-pipeline@v0.9.0 · 5738 in / 1338 out tokens · 33680 ms · 2026-05-23T20:56:22.567183+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages

  1. [1]

    Ameln, A

    Akcay, S., D. Ameln, A. Vaidya, B. Lakshmanan, N. Ahuja, and U. Genc (2022). Anomalib: A deep learning library for anomaly detection. In 2022 IEEE International Conference on Image Processing (ICIP) , pp.\ 1706--1710. IEEE

    work page 2022

  2. [2]

    Apostolopoulos, I. D. and M. A. Tzani (2023). Industrial object and defect recognition utilizing multilevel feature extraction from industrial scenes with deep learning approach. Journal of Ambient Intelligence and Humanized Computing\/ 14\/ (8), 10263--10276

    work page 2023

  3. [3]

    Fauser, D

    Bergmann, P., M. Fauser, D. Sattlegger, and C. Steger (2020). Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp.\ 4183--4192

    work page 2020

  4. [4]

    Chew, Y. Z. (2022). Anomaly detection for vision-based inspection . Ph.\ D. thesis, UTAR

    work page 2022

  5. [5]

    Prakash, E

    Christie, S., A. Prakash, E. Ashton, and V. Tanner (2023). Enhancing manual visual inspection process by using opencv ai and anomalib for defect detection in automotive assembly and manufacturing process. In International Electronic Packaging Technical Conference and Exhibition , Volume 87516, pp.\ V001T08A003. American Society of Mechanical Engineers

    work page 2023

  6. [6]

    Setkov, A

    Defard, T., A. Setkov, A. Loesch, and R. Audigier (2021). Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition , pp.\ 475--489. Springer

    work page 2021

  7. [7]

    Dewangan, D. K. and S. P. Sahu (2020). Deep learning-based speed bump detection model for intelligent vehicle system using raspberry pi. IEEE sensors journal\/ 21\/ (3), 3570--3578

    work page 2020

  8. [8]

    Pauchard, D

    D \"u rr, O., Y. Pauchard, D. Browarnik, R. Axthelm, and M. Loeser (2015). Deep learning on a raspberry pi for real time face recognition. In Eurographics (Posters) , pp.\ 11--12

    work page 2015

  9. [9]

    Fedorov, I

    Gorbachev, Y., M. Fedorov, I. Slavutin, A. Tugarev, M. Fatekhov, and Y. Tarkan (2019). Openvino deep learning workbench: Comprehensive analysis and tuning of neural networks inference. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops , pp.\ 0--0

    work page 2019

  10. [10]

    Ishizaka, and K

    Gudovskiy, D., S. Ishizaka, and K. Kozuka (2022). Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF winter conference on applications of computer vision , pp.\ 98--107

    work page 2022

  11. [11]

    Izumi, and L

    Hattori, K., T. Izumi, and L. Meng (2023). Defect detection of apples using patchcore. In 2023 International Conference on Advanced Mechatronic Systems (ICAMechS) , pp.\ 1--6. IEEE

    work page 2023

  12. [12]

    Hussain, M. (2023). Yolo-v1 to yolo-v8, the rise of yolo and its complementary nature toward digital manufacturing and industrial defect detection. Machines\/ 11\/ (7), 677

    work page 2023

  13. [13]

    Jha, S. B. and R. F. Babiceanu (2023). Deep cnn-based visual defect detection: Survey of current literature. Computers in Industry\/ 148 , 103911

    work page 2023

  14. [14]

    Jiang, X., G. Xie, J. Wang, Y. Liu, C. Wang, F. Zheng, and Y. Jin (2022). A survey of visual sensory anomaly detection. arXiv preprint arXiv:2202.07006\/

    work page arXiv 2022

  15. [15]

    Zhang, Y

    Jiang, Z., Y. Zhang, Y. Wang, J. Li, and X. Gao (2024). Fr-patchcore: An industrial anomaly detection method for improving generalization. Sensors\/ 24\/ (5), 1368

    work page 2024

  16. [16]

    Baik, and T

    Kim, D., S. Baik, and T. H. Kim (2023). Sanflow: Semantic-aware normalizing flow for anomaly detection. Advances in Neural Information Processing Systems\/ 36 , 75434--75454

    work page 2023

  17. [17]

    Kim, I., Y. Jeon, J. W. Kang, and J. Gwak (2023). Rag-padim: residual attention guided padim for defects segmentation in railway tracks. Journal of Electrical Engineering & Technology\/ 18\/ (2), 1429--1438

    work page 2023

  18. [18]

    Liu, J., G. Xie, J. Wang, S. Li, C. Wang, F. Zheng, and Y. Jin (2024). Deep industrial image anomaly detection: A survey. Machine Intelligence Research\/ 21\/ (1), 104--135

    work page 2024

  19. [19]

    Boumerdassi, A

    Mounir, T., S. Boumerdassi, A. Benhamada, A. Alla, and M. Kherarba (2020). Performance evaluation of basic image processing algorithms in cpu, gpu, raspberry pi and fpga. Int. J. Comput. Sci. Eng.(IJCSE)\/ 9\/ (4), 312--325

    work page 2020

  20. [20]

    Qurishee, M. A. (2019). Low-cost deep learning uav and raspberry pi solution to real time pavement condition assessment

    work page 2019

  21. [21]

    a hlte Beitr \

    Rippel, O. and D. Merhof (2023). Anomaly detection for automated visual inspection: A review. Bildverarbeitung in der Automation: Ausgew \"a hlte Beitr \"a ge des Jahreskolloquiums BVAu 2022\/ , 1--13

    work page 2023

  22. [22]

    Pemula, J

    Roth, K., L. Pemula, J. Zepeda, B. Sch \"o lkopf, T. Brox, and P. Gehler (2022). Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp.\ 14318--14328

    work page 2022

  23. [23]

    Russakovsky, O., J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. (2015). Imagenet large scale visual recognition challenge. International journal of computer vision\/ 115 , 211--252

    work page 2015

  24. [24]

    Tran, and O

    Santos, J., T. Tran, and O. Rippel (2023). Optimizing patchcore for few/many-shot anomaly detection. arXiv preprint arXiv:2307.10792\/

    work page arXiv 2023

  25. [25]

    S \"u zen, A. A., B. Duman, and B. S en (2020). Benchmark analysis of jetson tx2, jetson nano and raspberry pi using deep-cnn. In 2020 International Congress on Human-Computer Interaction, Optimization and Robotic Applications (HORA) , pp.\ 1--5. IEEE

    work page 2020

  26. [26]

    Tao, X., X. Gong, X. Zhang, S. Yan, and C. Adak (2022). Deep learning for unsupervised anomaly localization in industrial images: A survey. IEEE Transactions on Instrumentation and Measurement\/ 71 , 1--21

    work page 2022

  27. [27]

    Xie, G., J. Wang, J. Liu, J. Lyu, Y. Liu, C. Wang, F. Zheng, and Y. Jin (2024). Im-iad: Industrial image anomaly detection benchmark in manufacturing. IEEE Transactions on Cybernetics\/

    work page 2024

  28. [28]

    Yong, B. X. and A. Brintrup (2022). Bayesian autoencoders with uncertainty quantification: Towards trustworthy anomaly detection. Expert Systems with Applications\/ 209 , 118196

    work page 2022

  29. [30]

    Zheng, Y., X. Wang, Y. Qi, W. Li, and L. Wu (2022). Benchmarking unsupervised anomaly detection and localization. arXiv preprint arXiv:2205.14852\/

    work page arXiv 2022

  30. [31]

    Ziegler, T. (2023). Applications of AI on Resource-ConstrainedHardware with a focus on Anomaly Detection . Ph.\ D. thesis, Massachusetts Institute of Technology

    work page 2023

  31. [32]

    Verworner, M

    Zipfel, J., F. Verworner, M. Fischer, U. Wieland, M. Kraus, and P. Zschech (2023). Anomaly detection for industrial quality assurance: A comparative evaluation of unsupervised deep learning models. Computers & Industrial Engineering\/ 177 , 109045

    work page 2023

  32. [33]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

    work page

  33. [34]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

    work page

  34. [35]

    Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows,

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

    work page arXiv 2022

  35. [36]

    , " * write output.state after.block = add.period

    ENTRY address archive author booktitle chapter collaboration edition editor eid howpublished institution journal key lastchecked month note number numpages organization pages publisher school series title type url urldate volume year label extra.label sort.label INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.c...

    work page

  36. [37]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

    work page