CoChat
Sign Up
Cited by 0Industrial Vision Systems and Defect DetectionAdvanced Neural Network ApplicationsInfrastructure Maintenance and Monitoring

AI-Based Image Quality and Lens Defect Analysis in Autonomous Driving: A Framework with U-Net-Based Soiling Detection

Open access
O’g’li, Axmedov·2025·American Journal of Mechanics and Applications
DOI 10.11648/j.ajma.20251204.14
Open this paper in CoChat — free accountOpen PDF

Abstract

Ensuring reliable camera vision in autonomous driving systems requires continuous monitoring of image quality and lens integrity. External contaminants such as dust, raindrops, and mud, as well as permanent defects like cracks or scratches, can severely degrade visual perception and compromise safety-critical tasks such as lane detection, obstacle recognition, and path planning. This paper presents an AI-based framework that integrates image quality assessment (IQA) and lens defect analysis to enhance the robustness of camera-based perception systems in autonomous vehicles. Building on previous conceptual work in safety-aware lens defect detection, the proposed framework introduces a dual-layer architecture that combines real-time IQA monitoring with deep learning-based soiling segmentation. As an initial experimental validation, a U-Net model was trained on the WoodScape Soiling dataset to perform pixel-level detection of lens contamination. The model achieved an average Intersection-over-Union (IoU) of 0.6163, a Dice coefficient of 0.7626, and a recall of 0.9780, confirming its effectiveness in identifying soiled regions under diverse lighting and environmental conditions. Beyond the experiment, this framework outlines pathways for future integration of semantic segmentation, anomaly detection, and safety-driven decision policies aligned with ISO 26262 and ISO 21448 standards. By bridging conceptual modeling with experimental evidence, this study establishes a foundation for intelligent camera health monitoring and fault-tolerant perception in autonomous driving. The presented results demonstrate that AI-based image quality and defect assessment can significantly improve system reliability, supporting safer and more adaptive driving under real-world conditions.

AI snapshot

Written from the paper’s full text.

TL;DR

This paper proposes a conceptual AI-based framework that integrates image quality assessment, lens defect/soiling detection, and safety-aware decision logic for autonomous driving cameras. As an initial validation, a U-Net trained on a WoodScape Soiling subset achieved mean IoU 0.6163, Dice 0.7626, and recall 0.9780 for pixel-level soiling segmentation.

Key findings

  • The U-Net soiling segmentation model achieved test accuracy 0.6237, IoU 0.6163, Dice 0.7626, precision 0.6249, recall 0.9780, and specificity 0.0506.
  • Per-image IoU distribution had mean 0.6099, median 0.7112, standard deviation 0.3394, minimum 0.0306, and maximum 0.9941, indicating substantial variability by image complexity and lighting.
  • Threshold sensitivity showed optimal performance around 0.5; lower thresholds produced over-segmentation with recall 1.0000 and specificity 0.0000, while higher thresholds caused under-segmentation, with IoU 0.0066 at 0.6 and 0.0000 at 0.7.
  • The paper reports that high recall ensures minimal missed contamination regions, which aligns with safety-first design, though specificity remains low due to aggressive detection bias.

Methods

  • The framework is a four-layer conceptual architecture: image acquisition, perceptual analysis, decision/control aligned with ISO 26262 and ISO 21448, and maintenance/response.
  • Experimental validation used a custom subset of the WoodScape Soiling dataset with 499 training and 150 testing RGB images at 128 × 128 pixels, binary clean/contaminated masks, normalization, and augmentation including random rotation, flipping, and brightness scaling.
  • A U-Net was trained with Dice loss and Adam optimizer at learning rate 0.0001, batch size 8, up to 50 epochs with early stopping; training stopped at Epoch 14 with validation loss around 0.564.

Limitations

  • The paper states the model’s specificity remains low due to aggressive detection bias, and qualitative evaluation notes small false positives near edges.
  • The paper states the experiment is a proof-of-concept and that future work will address semi-supervised/unsupervised learning, multi-modal fusion, and real-time embedded deployment.
  • Evident from the text: only the U-Net soiling segmentation is experimentally validated; the IQA subsystem, safety decision logic, and permanent lens defects such as scratches or cracks are discussed conceptually but not experimentally tested.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.11648/j.ajma.20251204.14

Cites

14 of 14 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

3 of 6 cited sources checked · 1 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Existing studies mainly focused on classification or patch-based detection, providing limited spatial understanding of contamination.

    “However, these studies mainly focused on classification or patch-based detection, providing limited spatial understanding of contamination.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [1] Uřičář, M., Křížek, P., Sistu, G., & Yogamani, S. (2019, October). Soilingnet: Soiling detection on automotive surround-view cameras. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC) (pp. 67-72). IEEE. source

      The Formal definition (pp. 2–3) describes image-level classification and tile-level detection, supporting that characterization of SoilingNet. But it explicitly says tile-level detection provides higher spatial resolution, and Proposed CNN architecture (p. 4) reports an output configuration at image resolution. The claim’s suggestion that this study provides only limited spatial understanding therefore overstates what the source shows. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [2] A. Das, P. Křížek, G. Sistu, F. Bürger, S. Madasamy, M. Uřičár, V. R. Kumar, and S. Yogamani, “TiledSoilingNet: Tile-level Soiling Detection on Automotive Surround-View Cameras Using Coverage Metric,” in Proc. IEEE Intelligent Transportation Systems Conference (ITSC), Rhodes, Greece, Sept. 2020, pp. 1-6, https://doi.org/10.1109/ITSC45102.2020.9294677 source
  • Premisecitation

    Image restoration using inpainting or generative methods carries risks of hallucinated visual cues that may mislead the perception system.

    “Other works explored image restoration using inpainting or generative methods [3,4], though such approaches carry risks of hallucinated visual cues that may mislead the perception system.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [3] Soboleva, V., & Shipitko, O. (2021, December). Raindrops on windshield: Dataset and lightweight gradient-based detection algorithm. In 2021 IEEE Symposium Series on Computational Intelligence (SSCI) (pp. 1-7). IEEE. source
    • cites [4] S. Yogamani, M. Uricar, G. Sistu, et al., “WoodScape: A Multi-Task, Multi-Camera Fisheye Dataset for Autonomous Driving,” Proc. ICCV Workshops, 2019. source

      Section 2.2 discusses geometric undistortion of fisheye images, not restoration by inpainting or generative methods. The GDPR challenges section mentions GAN-generated face replacement for privacy, but does not address hallucinated visual cues or their effects on perception. Searches for inpainting, restoration, and hallucinated cues found no passage supporting the attributed point. (LLM, conf=0.75)

      Experimental — accuracy not yet measured
  • Findingown data

    The test set confusion matrix had TN = 47,520, FP = 891,471, FN = 33,390, and TP = 1,485,219.

    “The confusion matrix (TN = 47,520; FP = 891,471; FN = 33,390; TP = 1,485,219) indicates that the model tends to favor true positive detections while occasionally misclassifying clean regions as soiled pixels.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The binary ground truth masks accurately represent contamination in the images.
    • Rests on an assumption: The 128 × 128 pixel image resolution is sufficient for real-world soiling detection.
  • Findingown data

    The U-Net model achieved a specificity of 0.0506 on the test set.

    “Accuracy: 0.6237 IoU: 0.6163 Dice Score: 0.7626 Precision: 0.6249 Recall: 0.9780 Specificity: 0.0506”
    • Also stated in: § 4.7. Discussion · plain
      “Although the model's specificity remains low due to aggressive detection bias, its high recall ensures reliable identification of all potentially degraded regions.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The binary ground truth masks accurately represent contamination in the images.
    • Rests on an assumption: The 128 × 128 pixel image resolution is sufficient for real-world soiling detection.
  • Findingown data

    The U-Net model achieved an average Intersection-over-Union (IoU) of 0.6163 on the test set.

    “The model achieved an average Intersection-over-Union (IoU) of 0.6163, a Dice coefficient of 0.7626, and a recall of 0.9780”
    • Also stated in: § 4.3. Quantitative Results · plain
      “Accuracy: 0.6237 IoU: 0.6163 Dice Score: 0.7626 Precision: 0.6249 Recall: 0.9780 Specificity: 0.0506”
    • Also stated in: § 5. Conclusion and Future Work · plain
      “The model achieved a mean IoU of 0.6163, Dice score of 0.7626, and recall of 0.9780”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The binary ground truth masks accurately represent contamination in the images.
    • Rests on an assumption: The 128 × 128 pixel image resolution is sufficient for real-world soiling detection.
  • May combine two claims. 9
  • Classified as probably not a substantial claim. 6

  • May combine two claims.

    Findingown data

    The model successfully isolates large soiled regions and detects subtle blur or shadowed areas, though small false positives remain near edges.

    “The model successfully isolates large soiled regions and even detects subtle blur or shadowed areas, though small false positives remain near edges.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The binary ground truth masks accurately represent contamination in the images.
    • Rests on an assumption: The 128 × 128 pixel image resolution is sufficient for real-world soiling detection.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Early stopping occurred at Epoch 14, where validation loss stabilized around 0.564.

    “The training process began with an initial accuracy of approximately 0.50 and Dice coefficient near 0.54, showing gradual improvement before early stopping occurred at Epoch 14, where validation loss stabilized around 0.564.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The binary ground truth masks accurately represent contamination in the images.
    • Rests on an assumption: The 128 × 128 pixel image resolution is sufficient for real-world soiling detection.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    Training began with an initial accuracy of approximately 0.50 and a Dice coefficient near 0.54.

    “The training process began with an initial accuracy of approximately 0.50 and Dice coefficient near 0.54, showing gradual improvement before early stopping occurred at Epoch 14, where validation loss stabilized around 0.564.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The binary ground truth masks accurately represent contamination in the images.
    • Rests on an assumption: The 128 × 128 pixel image resolution is sufficient for real-world soiling detection.

  • May combine two claims.

    Findingown data

    Lower thresholds produced over-segmentation (high recall, low specificity), while higher thresholds caused under-segmentation (low recall, high specificity).

    “Lower thresholds produced over-segmentation (high recall, low specificity), while higher thresholds caused under-segmentation (low recall, high specificity).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The binary ground truth masks accurately represent contamination in the images.
    • Rests on an assumption: The 128 × 128 pixel image resolution is sufficient for real-world soiling detection.

  • May combine two claims.

    Interpretationown data

    The model tends to favor true positive detections while occasionally misclassifying clean regions as soiled pixels.

    “The confusion matrix (TN = 47,520; FP = 891,471; FN = 33,390; TP = 1,485,219) indicates that the model tends to favor true positive detections while occasionally misclassifying clean regions as soiled pixels.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Conventional cleaning mechanisms such as spray or wiper systems can remove transient soiling but are insufficient against structural lens defects.

    “Conventional cleaning mechanisms, such as spray or wiper systems, can remove transient soiling, but they are insufficient against structural lens defects that require intelligent diagnostic detection and system-level responses.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    The application of IQA techniques to autonomous driving remains limited, particularly in connecting image quality monitoring with safety-aware responses.

    “Yet, the application of IQA techniques to autonomous driving remains limited, particularly in connecting image quality monitoring with safety-aware responses.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The Dice coefficient curves show continuous improvement for both training and validation sets, confirming stable learning dynamics.

    “The corresponding Dice coefficient curves in Figure 3 show continuous improvement for both training and validation sets, confirming stable learning dynamics.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The training loss exhibits a consistent downward trend, whereas validation loss fluctuates moderately, indicating controlled generalization.

    “As depicted in Figure 2, the training loss exhibits a consistent downward trend, whereas validation loss fluctuates moderately, indicating controlled generalization.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Prior works by the authors provided theoretical and architectural foundations for camera health monitoring but did not include experimental validation at the pixel level.

    “These studies provided theoretical and architectural foundations for camera health monitoring but did not include experimental validation at the pixel level.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [5] Axmedov A. G. & Dadaxanov M. X. (2025). IMAGE QUALITY CHALLENGES IN AUTONOMOUS DRIVING: A CONCEPTUAL FRAMEWORK FOR CAMERA LENS DEFECT DETECTION. Development Of Science, 9(2), pp. 69-79.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [6] Axmedov Abdulazizxon Ganijon O'g'li, & Dadaxanov Musoxon Xoshimxonovich. (2025). ENHANCING AUTONOMOUS DRIVING SYSTEMS WITH AI-BASED IMAGE QUALITY ASSESSMENT FOR LENS DEFECT DETECTION. PORTUGAL-SCIENTIFIC REVIEW OF THE PROBLEMS AND PROSPECTS OF MODERN SCIENCE AND EDUCATION, 1(3), 38-43.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The proposed framework establishes a functional safety bridge between perception quality degradation and operational control through the Severity-Criticality-Recoverability (S-C-R) model.

    “Unlike prior works that treated soiling detection and IQA separately, the proposed framework establishes a functional safety bridge between perception quality degradation and operational control through the Severity-Criticality-Recoverability (S-C-R) model.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The integration of IQA and U-Net soiling detection into a unified framework is feasible without further validation.

  • May combine two claims.

    Interpretationown data

    The model's high recall ensures minimal missed contamination regions, an essential characteristic for safety-critical applications.

    “Nevertheless, high recall (0.9780) ensures minimal missed contamination regions-an essential characteristic for safety-critical applications.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The U-Net model's performance on the test set generalizes to real-world autonomous driving scenarios.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The model's conservative behavior aligns with safety-first system design, where detecting contamination is prioritized over missing defects.

    “This conservative behavior aligns with safety-first system design, where detecting contamination is prioritized over missing defects.”
    Quote in the paperSubstantial claim
13 of 13

  • The paper relies on this without support.

    The WoodScape Soiling dataset (and the custom subset used) is representative of real-world camera soiling conditions in autonomous driving.

    “The experimental validation of the proposed framework was conducted using a custom subset of the WoodScape Soiling dataset, designed for pixel-level detection of lens contamination.”

  • The paper relies on this without support.

    The U-Net model's performance on the test set generalizes to real-world autonomous driving scenarios.

    “The model achieved an average Intersection-over-Union (IoU) of 0.6163, a Dice coefficient of 0.7626, and a recall of 0.9780, confirming its effectiveness in identifying soiled regions under diverse lighting and environmental conditions.”

  • The paper relies on this without support.

    The binary ground truth masks accurately represent contamination in the images.

    “Ground truth segmentation masks contained binary labels representing clean and contaminated regions”

  • The paper relies on this without support.

    The 128 × 128 pixel image resolution is sufficient for real-world soiling detection.

    “each with spatial dimensions of 128 × 128 pixels”

  • The paper relies on this without support.

    The integration of IQA and U-Net soiling detection into a unified framework is feasible without further validation.

    “The proposed framework introduces a dual-layer architecture that combines real-time IQA monitoring with deep learning-based soiling segmentation.”
5 of 5

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • SoilingNet is described as both patch-based and pixel-level

    Minor

    The introduction groups SoilingNet with studies said to focus on classification or patch-based detection with limited spatial understanding, while Section 2.1 explicitly describes it as a pixel-level network.

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 2

  1. Axmedov A. G. & Dadaxanov M. X. (2025). IMAGE QUALITY CHALLENGES IN AUTONOMOUS DRIVING: A CONCEPTUAL FRAMEWORK FOR CAMERA LENS DEFECT DETECTION. Development Of Science, 9(2), pp. 69-79.
  2. Axmedov Abdulazizxon Ganijon O'g'li, & Dadaxanov Musoxon Xoshimxonovich. (2025). ENHANCING AUTONOMOUS DRIVING SYSTEMS WITH AI-BASED IMAGE QUALITY ASSESSMENT FOR LENS DEFECT DETECTION. PORTUGAL-SCIENTIFIC REVIEW OF THE PROBLEMS AND PROSPECTS OF MODERN SCIENCE AND EDUCATION, 1(3), 38-43.

Authors
O’g’li, A.
Affiliations
Namangan State University (UZ)
Corresponding author
O’g’li, Axmedov
ORCID
O’g’li, Axmedov
Venue
American Journal of Mechanics and Applications
Publisher
Science Publishing Group
Published
2025-12-17
Volume
12
Issue
4
Pages
93-101
ISSN
  • 2376-6115 (ISSN-L)
  • 2376-6131
DOI
10.11648/j.ajma.20251204.14
OpenAlex
W7116763411
Semantic Scholar
6839074b1862b9823bb0f74d9e80496f53cb81a7
Cited by
0
References
13
FWCI
0.00
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
American Journal of Mechanics and Applications (publishedVersion, cc-by)
Topics
  • Physical Sciences › Engineering › Industrial and Manufacturing Engineering › Industrial Vision Systems and Defect Detection
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Engineering › Civil and Structural Engineering › Infrastructure Maintenance and Monitoring
Keywords
image quality assessment, autonomous driving, U-Net, ISO 26262, ISO 21448, real-time monitoring

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review