Pith. sign in

REVIEW 3 major objections 6 minor 32 references

A Fuzzy Rule-based Neuro-Symbolic Approach for Pipe Severity Prediction in Sewer Networks

T0 review · 3 major / 6 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Decoupling neural defect recognition from fixed fuzzy rules improves sewer pipe severity prediction and makes every decision traceable from CODE degrees to class evidence.

desk verdict Solid applied modular neuro-symbolic sewer severity system with real gains and honest traces; the unvalidated LLM severity labels and tiny rare classes are the main soft spots, not the architecture. read the letter →

arxiv 2607.28481 v1 pith:X7MOQ7TX submitted 2026-07-30 cs.AI

Ngoc Thai Le , Thanh Ma , Umberto Straccia This is my paper · ORCID
classification cs.AI
keywords FuzzyRulesNeuro-SymbolicMachineLearningSewerNetworksMultilabelDefectCODEsDecisionTreeInterpretableSeverityAssessment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Direct image-to-severity models for sewer pipes hide how visual defects become maintenance priorities. This paper inserts standardized inspection CODEs as an intermediate semantic layer: a Swin Transformer predicts graded degrees for 14 multilabel CODEs from each image, then a fixed base of 19 IF–THEN rules—extracted once from a decision tree on ground-truth CODEs—scores severity with fuzzy t-norm activation, rule confidence, and s-norm aggregation. On a 3,244-image, five-class imbalanced French inspection set whose severity targets came from five-LLM consensus on inspector notes, the approach raises accuracy, balanced accuracy, Macro F1, and MCC by about 18%, 12%, 23%, and 17% over the same backbone used as a pure image classifier. Soft-degree Product and Hamacher operators win validation selection. The practical payoff is an image-only deployment path that still exposes which CODEs fired which rules and how much evidence each severity class received.

What carries the argument

The modular perception–reasoning split: Swin Transformer CODE degrees → condition memberships (μ or 1−μ) → t-norm rule activation × leaf confidence → s-norm class evidence. The J48-extracted rule base is frozen after training; soft-degree fuzzy evaluation is the selected interface.

What would settle it

Have independent human inspectors re-score the held-out test images under the same guidelines, blind to LLM labels and CODEs; if the Multilabel+Rule advantage over the image-only baseline vanishes or reverses under those human severity labels, the central performance claim fails.

Watch

Extended reading notes

Core claim

Predicting multilabel defect CODE degrees from sewer images and feeding them into a fixed 19-rule fuzzy reasoner (Product, Łukasiewicz, or Hamacher) outperforms direct image-to-severity classification on class-balanced metrics while yielding a full numerical trace from CODE conditions through rule supports to aggregated severity evidence.

Load-bearing premise

Severity labels built only by five-LLM consensus on free-text inspector notes—without expert review and with CODEs withheld—are accurate enough that the reported gains reflect real severity reasoning rather than labeling artifacts, despite very few severe test cases.

Editorial extensions

If this is right

  • A severity call can be audited by reading predicted CODE degrees, activated rules, confidences, and class evidence instead of a single black-box score.
  • Stronger CODE predictors can be swapped in without retraining the symbolic rule layer.
  • Soft-degree fuzzy aggregation can outperform crisp CODE thresholds when visual evidence is graded.
  • Oracle CODE+Rule results show that better CODE recognition alone would close much of the remaining gap to the rule-base ceiling.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same CODE-bottleneck plus fixed fuzzy rules pattern could transfer to other coded infrastructure inspections (bridges, pavements, tunnels) that already standardize defects.
  • If rare high-severity classes stay weak even with perfect CODEs, richer still-traceable rule extractors may be needed without abandoning the single-path explanation style.
  • Multi-LLM consensus on inspector notes may be a workable way to create severity targets when expert re-annotation is scarce, but still needs human spot-checks before maintenance decisions rest on it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes a modular neuro-symbolic pipeline for sewer-pipe severity prediction that decouples Swin-Transformer multilabel prediction of 14 inspection CODE degrees from a fixed 19-rule fuzzy reasoner extracted from a single J48 tree on ground-truth CODEs and severity labels. Severity targets are obtained by five-LLM consensus over inspector free-text notes (CODEs excluded). On a fixed 3,244-image split, validation-selected soft-degree Product/Hamacher inference improves accuracy, balanced accuracy, Macro F1, and MCC by about 17.9%, 12.2%, 23.0%, and 17.3% over a same-backbone image-only classifier (Table 6), while exposing full traces from CODE degrees through rule supports to class evidence. An Oracle CODE+Rule setting localizes much of the remaining error to CODE prediction.

Significance. If the reported gains hold under reliable severity targets and stronger statistical controls, the work is a useful contribution to interpretable infrastructure inspection: it makes the practical CODE→severity workflow explicit, supplies a compact fixed rule base with fuzzy t-/s-norm aggregation, and demonstrates that intermediate defect semantics can beat direct image-to-severity classification while remaining deployable image-only. Strengths include clear modular separation (no joint rule-loss training), leakage checks, a well-motivated Oracle diagnostic, multi-operator robustness checks, and qualitative numerical traces (Tables 8–9). The main scientific value is transparency and engineering auditability rather than a new end-to-end learning paradigm.

major comments (3)
  1. [Section 4.1; Table 6] Section 4.1 and Table 6: All headline relative gains are measured exclusively against five-LLM consensus severity labels derived from inspector notes, with no domain-expert audit and with CODEs deliberately excluded. The paper does not report inter-LLM agreement, human–LLM agreement, or any expert subsample validation. Because both the image-only baseline and Multilabel+Rule are scored on the same unvalidated targets—and rules are fit to those same labels on train—the central claim that the lift reflects better severity reasoning (vs. better matching of note-pipeline artifacts) is not yet secured. At minimum, provide expert review on a stratified subsample and agreement statistics; without that, treat Table 6 gains as provisional.
  2. [Table 4; Appendix B; Abstract] Table 4 and Appendix B (Table 10): Test support for minority severities is extremely thin (class 4: 13 images; class 5: 4). Under Product soft-degree, class-4 recall is 7.7% while class-1 recall is 93.8%. Aggregate Macro F1/MCC improvements can therefore be driven by majority-class behavior and a handful of minority cases. The manuscript should report per-class precision/recall with uncertainty (bootstrap or repeated splits), and temper Abstract/§5 claims that imply balanced severity assessment when class 4 remains essentially unrecognized in the deployable setting.
  3. [Section 5.2–5.3; Table 7; Abstract] Section 5.2–5.3 and Table 7: Evaluation uses a single fixed split; soft-degree beats the best threshold by only 0.0015–0.0033 Macro F1 on validation, yet is locked in for all test claims. No confidence intervals, repeated splits, or cross-validation are reported for either CODE prediction or severity metrics. Given the acknowledged split-dependence and free choices (J48 -C/-M, interface mode, LLM aggregation), the load-bearing relative percentages in the Abstract need uncertainty quantification or multi-split confirmation before they can be treated as stable effect sizes.
minor comments (6)
  1. [Section 3.4] Typo in §3.4: “Spcifically” → “Specifically”.
  2. [Section 5.5; Table 6] §5.5 cites Oracle MCC as 0.7341 while Table 6 reports 0.7351; align the numbers.
  3. [Figure 1] Figure 1 is helpful but the online path’s fuzzy operators are only named in text; a small annotation of t-norm/s-norm and Conf_r on the figure would aid readers.
  4. [Appendix A] Appendix A lists 19 rules clearly; adding leaf (N_r/E_r) or Conf_r next to each rule would make the confidence weighting fully auditable without re-deriving from Weka.
  5. [Section 2] Related work could briefly contrast with recent concept-bottleneck and neuro-symbolic sewer/condition papers beyond the cited CBMs and NeurASP/LTN line, if space allows.
  6. arXiv date stamp in the PDF header reads “30 Jul 2026”; verify metadata.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: severity targets, CODE→rule extraction, and image-only baseline are independently constructed; reported lifts are ordinary held-out comparisons.

full rationale

The derivation chain does not collapse inputs into claimed outputs by construction. Severity labels y are produced offline from inspector free-text notes via five-LLM consensus with CODEs deliberately excluded (Section 4.1), so the target is not defined from the intermediate CODE layer. The symbolic rule base is a fixed J48 tree fit only on train ground-truth CODEs and those severity labels; root-to-leaf paths become 19 IF–THEN rules with leaf confidences Conf_r = (N_r−E_r)/N_r—standard empirical leaf purity, not a re-labeling of the test quantity. At deployment the neural module maps image→μ (14 CODE degrees) and the fixed rules map μ→ŷ via chosen t-/s-norms; the image-only baseline is a separate same-backbone multiclass classifier trained directly on images→severity. Both are scored on the held-out test split against the same external y, with interface selection confined to validation Macro F1. Oracle CODE+Rule is explicitly diagnostic (perfect CODE input), not a deployable “prediction.” Self-citations (fuzzy t-norm background, Fuzzy OWL-Boost/PN-OWL) supply operator definitions only and do not force the Table 6 lifts. Residual concerns about LLM-label validity or shared inspection practice are label-quality/correctness issues, not circular reduction of the claimed improvements.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The claim rests on standard multilabel CV and fuzzy-logic operators, a domain workflow assumption that CODE combinations determine severity, proprietary labeled images, LLM-consensus severity construction, and a single pruned J48 extraction treated as the symbolic knowledge base. No new physical entities; free choices are classical ML hyperparameters and the validation-selected soft interface.

free parameters (5)
  • J48 pruning confidence factor and min leaf size = C=0.25, M=2
    weka J48 -C 0.25 -M 2 fixes rule granularity and the 19-rule base shape.
  • Swin multilabel training hyperparameters = lr=1e-4, wd=1e-4, epochs=30, batch=64, weights clipped [1,30]
    AdamW lr, weight decay, 30 epochs, batch 64, positive CODE weights clipped to [1,30], ReduceLROnPlateau on val mAP.
  • CODE-interface mode per fuzzy operator = soft-degree (all three operators)
    Chosen on validation Macro F1 among soft-degree and τ∈{0.1..0.9}; soft-degree selected for P/L/H.
  • Multi-LLM severity aggregation = Q=5, round(mean), clip to [1,5]
    Five models’ integer scores averaged, rounded, clipped to {1..5}; defines all y labels.
  • Rule confidence Conf_r = (N_r - E_r)/N_r = per-rule empirical leaf purity
    Leaf class purity from the training DT becomes the multiplicative weight in Supp_ir.
assumptions (6)
  • standard math Standard multilabel BCE training and sigmoid CODE degrees in [0,1] are valid graded presence evidence for symbolic conditions.
    Section 3.3; degrees used directly as condition memberships.
  • standard math Product, Łukasiewicz, and Hamacher t-norm/s-norm pairs are appropriate conjunction/disjunction semantics for rule activation and class aggregation.
    Section 3.5 definitions and experimental sweep.
  • domain assumption Inspector CODEs form a sufficient semantic bottleneck: severity is a function of CODE presence/absence patterns capturable by a shallow DT rule base.
    Core hypothesis in Introduction and Section 3.1; motivates decoupling perception from reasoning.
  • ad hoc to paper Five-LLM consensus on French inspector notes yields severity labels accurate enough for training and metric claims, without using CODE annotations.
    Section 4.1 protocol; authors note need for expert validation in limitations.
  • ad hoc to paper A single pruned J48 tree on train CODEs+severity is an adequate fixed knowledge base (19 rules, 14 CODEs) for deployment reasoning.
    Section 3.4; RIPPER and ensembles deferred to future work.
  • domain assumption Fixed train/val/test split with path and (source file, photo id) checks implies no material leakage for reported test metrics.
    Section 4.2 leakage protocol.
invented entities (2)
  • Fixed 19-rule fuzzy severity knowledge base over 14 retained CODEs
    purpose: Interpretable symbolic mapping from graded CODEs to five-class severity evidence.
    Extracted artifact of this dataset/DT, not a general law; listed fully in Appendix A.
  • Multi-LLM consensus severity labeling protocol for inspector notes
    purpose: Create y labels without consuming CODE annotations reserved for rules and multilabel training.
    Paper-specific label factory; no external benchmark of agreement with human severity experts is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Fuzzy Rule-based Neuro-Symbolic Approach for Pipe Severity Prediction in Sewer Networks." pith.science (2026). https://pith.science/paper/X7MOQ7TX

@misc{pith2026260728481,
  author       = {Pith},
  title        = {Pith review of: A Fuzzy Rule-based Neuro-Symbolic Approach for Pipe Severity Prediction in Sewer Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X7MOQ7TX}},
  note         = {Machine review of arXiv:2607.28481}
}
read the original abstract

Standard automated sewer pipe severity assessment relies on direct image classification, creating a "black box" where the link between visual defects and final severity scores remains implicit. This study introduces a modular, fuzzy rule-based neuro-symbolic framework that bridges this gap by decoupling neural perception from symbolic reasoning. The perception module utilizes a Swin Transformer to predict 14 multilabel inspection CODE degrees directly from images. For reasoning, a DT, specifically Weka's J48, algorithm is trained on ground-truth CODEs and severity labels, and its paths are converted into 19 fixed IF--THEN rules. Inference operates via fuzzy logic: t-norm activations from CODE conditions are weighted by rule confidence and combined with corresponding s-norms to produce interpretable class evidence. We assessed Product, {\L}ukasiewicz, and Hamacher operator pairs using a dataset of 3,244 images spanning five highly imbalanced severity classes. Ground-truth labels were robustly generated via consensus from five independent large language models analyzing original inspector notes. Our results show an improvement of accuracy, balanced accuracy, Macro F1 and MCC by 17.9%, 12.2%, 23.0%, and 17.3%, respectively, over image-only based classification. Overall, the framework combines competitive class-balanced performance with traceable reasoning from predicted CODE degrees to rule supports and severity evidence.

Figures

Figures reproduced from arXiv: 2607.28481 by the authors.

Figure 1
Figure 1. Overall framework of the proposed fuzzy rule-based neuro-symbolic [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 3 linked inside Pith

  1. [1]

    Modeling the structural deterioration of urban drainage pipes: the state-of-the-art in statistical methods.Urban Water Journal, 7(1):47–59, 2010

    EV Ana and Willy Bauwens. Modeling the structural deterioration of urban drainage pipes: the state-of-the-art in statistical methods.Urban Water Journal, 7(1):47–59, 2010

    2010

  2. [2]

    Logic tensor networks.Artificial Intelligence, 303:103649, 2022

    Samy Badreddine, Artur d’Avila Garcez, Luciano Serafini, and Michael Spranger. Logic tensor networks.Artificial Intelligence, 303:103649, 2022

    2022

  3. [3]

    Pn-owl: A two stage algorithm to learn fuzzy concept inclusions from owl ontologies

    Franco Alberto Cardillo, Franca Debole, and Umberto Straccia. Pn-owl: A two stage algorithm to learn fuzzy concept inclusions from owl ontologies. Fuzzy Sets and Systems, 490(109048), 2024

    2024

  4. [4]

    Fuzzy owl-boost: Learning fuzzy concept inclusions via real-valued boosting.Fuzzy Sets and Systems, 438:164–186, 2022

    Franco Alberto Cardillo and Umberto Straccia. Fuzzy owl-boost: Learning fuzzy concept inclusions via real-valued boosting.Fuzzy Sets and Systems, 438:164–186, 2022. 27

    2022

  5. [5]

    Neuro-fuzzy approaches for san- itary sewer pipeline condition assessment.Journal of Computing in Civil engineering, 15(1):4–14, 2001

    Myung Jin Chae and Dulcy M Abraham. Neuro-fuzzy approaches for san- itary sewer pipeline condition assessment.Journal of Computing in Civil engineering, 15(1):4–14, 2001

    2001

  6. [6]

    Fast effective rule induction

    William W Cohen. Fast effective rule induction. InProceedings of the Twelfth International Conference on Machine Learning (ICML), pages 115–123. Morgan Kaufmann, 1995

    1995

  7. [7]

    Neural logic machines.arXiv preprint arXiv:1904.11694, 2019

    Honghua Dong, Jiayuan Mao, Tian Lin, Chong Wang, Lihong Li, and Denny Zhou. Neural logic machines.arXiv preprint arXiv:1904.11694, 2019

    arXiv 1904

  8. [8]

    Approaches to sewer maintenance: A review.Urban Water, 2:343–356, 12 2000

    R.A Fenner. Approaches to sewer maintenance: A review.Urban Water, 2:343–356, 12 2000

    2000

Show all 32 references
  1. [9]

    Un- derground sewer pipe condition assessment based on convolutional neural networks.Automation in Construction, 106:102849, 2019

    Syed Ibrahim Hassan, L Minh Dang, Irfan Mehmood, Suhyeon Im, Changho Choi, Jaemo Kang, Young-Soo Park, and Hyeonjoon Moon. Un- derground sewer pipe condition assessment based on convolutional neural networks.Automation in Construction, 106:102849, 2019

    2019

  2. [10]

    Multi-task classification of sewer pipe defects and properties using a cross-task graph neural network decoder

    Joakim Bruslund Haurum, Meysam Madadi, Sergio Escalera, and Thomas B Moeslund. Multi-task classification of sewer pipe defects and properties using a cross-task graph neural network decoder. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pa...

    2022

  3. [11]

    A survey on image- based automation of cctv and sset sewer inspections.Automation in Con- struction, 111:103061, 2020

    Joakim Bruslund Haurum and Thomas B Moeslund. A survey on image- based automation of cctv and sset sewer inspections.Automation in Con- struction, 111:103061, 2020

    2020

  4. [12]

    George J. Klir. Where do we stand on measures of uncertainty, ambiguity, fuzziness, and the like?Fuzzy Sets Syst., 24(2):141–160, 1987

    1987

  5. [13]

    Klir and Bo Yuan.Fuzzy sets and fuzzy logic: theory and appli- cations

    George J. Klir and Bo Yuan.Fuzzy sets and fuzzy logic: theory and appli- cations. Prentice-Hall, Inc., Upper Saddle River, NJ, USA, 1995

    1995

  6. [14]

    Concept bottleneck models

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338–5348. PMLR, 2020

    2020

  7. [15]

    Automated defect classification in sewer closed circuit television inspections using deep convolutional neural networks.Automa- tion in Construction, 91:273–283, 2018

    Srinath S Kumar, Dulcy M Abraham, Mohammad R Jahanshahi, Tom Ise- ley, and Justin Starr. Automated defect classification in sewer closed circuit television inspections using deep convolutional neural networks.Automa- tion in Construction, 91:273–283, 2018

    2018

  8. [16]

    Sewer damage detection from imbalanced cctv inspection data using deep convolutional neural networks with hierarchical classification.Automation in Construction, 101:199–208, 2019

    Duanshun Li, Anran Cong, and Shuai Guo. Sewer damage detection from imbalanced cctv inspection data using deep convolutional neural networks with hierarchical classification.Automation in Construction, 101:199–208, 2019

    2019

  9. [17]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF international con- ference on computer vision, pages 10012–10022, 2021. 28

    2021

  10. [18]

    A model of mul- timedia information retrieval.Journal of the ACM, 48(5):909–970, 2001

    Carlo Meghini, Fabrizio Sebastiani, and Umberto Straccia. A model of mul- timedia information retrieval.Journal of the ACM, 48(5):909–970, 2001

    2001

  11. [19]

    Mitchell.Machine Learning

    Tom M. Mitchell.Machine Learning. McGraw-Hill series in computer science. McGraw-Hill, 1 edition, 1997

    1997

  12. [20]

    Ross Quinlan

    J. Ross Quinlan. Simplifying decision trees.International Journal of Man- Machine Studies, 27(5):221–234, 1987

    1987

  13. [21]

    Ross Quinlan.C4.5: Programs for Machine Learning

    J. Ross Quinlan.C4.5: Programs for Machine Learning. Morgan Kauf- mann, 1993

    1993

  14. [22]

    Neuro-symbolic artificial intelligence: Current trends.arXiv preprint arXiv:2105.05330, 2021

    Md Kamruzzaman Sarker, Lu Zhou, Aaron Eberhart, and Pascal Hit- zler. Neuro-symbolic artificial intelligence: Current trends.arXiv preprint arXiv:2105.05330, 2021

    2021 arXiv

  15. [23]

    Neuro-fuzzy network for the classi- fication of buried pipe defects.Automation in Construction, 15(1):73–83, 2006

    Sunil K Sinha and Paul W Fieguth. Neuro-fuzzy network for the classi- fication of buried pipe defects.Automation in Construction, 15(1):73–83, 2006

    2006

  16. [24]

    Classification of underground pipe scanned images using feature extraction and neuro-fuzzy algorithm.IEEE Transactions on Neural Networks, 13(2):393–401, 2002

    Sunil K Sinha and Fakhri Karray. Classification of underground pipe scanned images using feature extraction and neuro-fuzzy algorithm.IEEE Transactions on Neural Networks, 13(2):393–401, 2002

    2002

  17. [25]

    Managing uncertainty and vagueness in description log- ics, logic programs and description logic programs

    Umberto Straccia. Managing uncertainty and vagueness in description log- ics, logic programs and description logic programs. InReasoning Web, 4th International Summer School, Tutorial Lectures, volume 5224 ofLecture Notes in Computer Science, pages 54–103. Springer Verlag, 2008

    2008

  18. [26]

    CRC Studies in Informatics Series

    Umberto Straccia.Foundations of Fuzzy Logic and Semantic Web Lan- guages. CRC Studies in Informatics Series. Chapman & Hall, 2013

    2013

  19. [27]

    Sewer asset management–state of the art and research needs.Urban Water Journal, 16(9):662–675, 2019

    Franz Tscheikner-Gratl, Nicolas Caradot, Fr´ ed´ eric Cherqui, Joao P Leit˜ ao, Mehdi Ahmadi, Jeroen G Langeveld, Yves Le Gat, Lisa Scholten, Bardia Roghani, Juan Pablo Rodr ´ ıguez, et al. Sewer asset management–state of the art and research needs.Urban Water Journal, 16(9):6...

    2019

  20. [28]

    J. D. Ullman.Principles of Database and Knowledge Base Systems, volume 1,2. Computer Science Press, Potomac, Maryland, 1989

    1989

  21. [29]

    Mingzhu Wang, Han Luo, and Jack CP Cheng. Towards an automated con- dition assessment framework of underground sewer pipes based on closed- circuit television (cctv) images.Tunnelling and Underground Space Tech- nology, 110:103840, 2021

    2021

  22. [30]

    Morgan Kaufmann, 3rd edition, 2011

    Ian H Witten, Eibe Frank, and Mark A Hall.Data Mining: Practical Machine Learning Tools and Techniques. Morgan Kaufmann, 3rd edition, 2011

    2011

  23. [31]

    Neurasp: Embracing neural networks into answer set programming.arXiv preprint arXiv:2307.07700, 2023

    Zhun Yang, Adam Ishay, and Joohyung Lee. Neurasp: Embracing neural networks into answer set programming.arXiv preprint arXiv:2307.07700, 2023. 29

    2023 arXiv

  24. [32]

    Automatic sewer defect detection and severity quantification based on pixel-level semantic segmentation.Tunnelling and Underground Space Technology, 123:104403, 2022

    Qianqian Zhou, Zuxiang Situ, Shuai Teng, Hanlin Liu, Weifeng Chen, and Gongfa Chen. Automatic sewer defect detection and severity quantification based on pixel-level semantic segmentation.Tunnelling and Underground Space Technology, 123:104403, 2022. 30 Table 10: Row-normalize...

    2022

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.