Pith. sign in

REVIEW 4 major objections 6 minor 31 references

Improved YOLOv7 model for insulator defect detection

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Three modifications to YOLOv7 lift insulator-defect detection mAP from 93.3% to 94.9% on a private drone-image dataset.

desk verdict Routine YOLOv7 tweak with an ablation table that contradicts its own text; the 1.6-point mAP gain is plausible but unverified. read the letter →

arxiv 2502.07179 v1 pith:NGBIWS7L submitted 2025-02-11 cs.CV cs.AI

Zhenyue Wang , Guowu Yuan , Hao Zhou , Yi Ma , Yutang Ma , Dong Chen This is my paper · ORCID
classification cs.CVcs.AI
keywords insulatordefectdetectionYOLOv7receptivefieldblockcoordinateattentionWIoUlossdroneinspectionobject
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

The paper tries to show that a standard YOLOv7 detector can be made substantially better at finding damaged power-line insulators in drone images by swapping in three modules: a receptive-field block in place of the SPPCSPC feature-extraction module, a coordinate-attention mechanism in the detection head, and a wise-IoU loss for training. On the authors' private 3000-image dataset of normal, self-exploded, and partially damaged insulators, the combined model reports mAP_0.5 rising from 93.3% to 94.9%, mAP_0.5:0.95 from 73.5% to 75.1%, precision from 92% to 93.3%, and recall from 91.1% to 92.1%, while using 3.2 million fewer parameters and less compute. If the improvement is real, it would make automated drone inspection of transmission lines more reliable across insulator types and damage shapes, which is what the paper argues.

What carries the argument

RFB (receptive field block) replaces SPPCSPC to enlarge effective receptive fields via multi-branch convolutions with dilated kernels, improving feature extraction for varied defect shapes. CA (coordinate attention) is inserted into the head to encode position information into channel attention and better localize small defect regions. WIoU (wise intersection over union) loss, specifically WIoUv3, dynamically down-weights low-quality anchor boxes during training so the model focuses on regular-quality samples. The combination carries the claimed accuracy gains; the paper argues each addresses one of the three stated difficulties (diverse materials and colors, complex backgrounds, and small or diverse defects).

What would settle it

Train the baseline and the proposed model on the same data with five or more random seeds; if the seed-to-seed variation in mAP_0.5 exceeds the reported 1.6-point gain, the improvement is indistinguishable from noise. Alternatively, test both models on an independent public insulator-defect dataset and check whether the gain reproduces.

Watch

Extended reading notes

Core claim

The central discovery is that each of the three proposed components contributes to better multi-type insulator defect detection, and that their combination outperforms the baseline YOLOv7 on all reported metrics. Replacing SPPCSPC with RFB alone improves mAP_0.5 by 0.8 points and mAP_0.5:0.95 by 1.1 points while cutting parameters; adding CA raises mAP_0.5 further and improves localization of small self-explosion defects; adding WIoU loss brings recall up to 92.1%. The paper also shows that CA outperforms CBAM, ECA, GAM, and SimAM when inserted at the same location, and that WIoUv3 converges faster than CIoU on their data.

Load-bearing premise

The load-bearing premise is that the reported 1.6-point mAP improvement is a real effect rather than training noise, since Tables 1 and 3 report single runs without seeds or error bars on one private 3000-image dataset.

Editorial extensions

If this is right

  • Automated grid inspection can detect multiple defect types (normal, self-explosion, partial damage) in one pass with higher recall, reducing missed defects.
  • The 3.2 million parameter reduction and 2.5 GFLOPS saving mean the improved model is cheaper to deploy than baseline YOLOv7, though the paper notes it is still not fast enough for embedded devices.
  • Per-category gains, such as self-explosion mAP_0.5 rising from 89.1% to 91.5%, suggest the method improves detection across defect classes, not only on average.
  • The ablation order indicates RFB alone provides most of the parameter savings, while CA and WIoU contribute accuracy, so future YOLO variants could adopt these modules independently.

Reading between the lines

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

  • If the single-run results hold across random seeds, similar gains might transfer to other one-stage detectors such as YOLOv5 and YOLOv8, since RFB, CA, and WIoU are architecture-agnostic modules.
  • The private dataset and the absence of seed-averaged results make the 1.6-point gain hard to verify; an independent evaluation on a public insulator-defect benchmark would test whether the improvement is dataset-specific.
  • The 2.81 ms speed improvement is partly a byproduct of parameter reduction, but the paper's own conclusion concedes the detector is still too slow for embedded deployment, so the practical operational gain is limited to offline or drone-based review.
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

4 major / 6 minor

Summary. The paper proposes an improved YOLOv7 model for detecting multiple types of insulator defects (normal, self-explosion, and partial damage) in drone-acquired power-line images. Three modifications are introduced: replacing the SPPCSPC module with a receptive field block (RFB), adding a coordinate attention (CA) mechanism to the head, and using the WIoUv3 loss function. On a private dataset of 3000 images from Yunnan Power Grid, the authors report that the combined model improves mAP_0.5 from 93.3% to 94.9%, mAP_0.5:0.95 from 73.5% to 75.1%, precision from 92.0% to 93.3%, and recall from 91.1% to 92.1%, while reducing parameters, GFLOPs, and inference time. Additional comparisons against other attention mechanisms, loss functions, and detectors (Faster R-CNN, YOLOv5m, YOLOv6m, YOLOv8m) are provided.

Significance. If the reported gains are reproducible, the work is a modest but practically relevant engineering contribution: it combines three known components (RFB, CA, WIoU) with YOLOv7 for a challenging multi-class defect-detection task. The use of standard, externally validated modules and the reporting of speed/parameter metrics are strengths, as is the coverage of three defect types in one dataset. However, the experimental evidence is the weakest part: all results come from single training runs without seeds or error bars, the dataset is private, and the ablation tables contain internal inconsistencies that undermine the component-wise attribution of gains. These issues must be addressed before the main claim can be accepted.

major comments (4)
  1. [Section 4.3.3, Table 1] The text states that replacing SPPCSPC with the RFB module improves mAP_0.5 by 0.8%, mAP_0.5:0.95 by 1.1%, precision by 1.4%, and reduces parameters by 3.3M, computation by 2.6 GFLOPS, and speed by 5.7ms. However, the RFB-only row in Table 1 shows parameters 37.7M (higher than the 37.2M baseline), GFLOPs 105.2 (similar to baseline), and speed 38.4ms (slower than the 36.9ms baseline). The claimed deltas match the RFB+CA row (33.9M, 102.5, 31.2ms), not the RFB-only row. This is a load-bearing inconsistency because the ablation is the only evidence attributing each improvement to a specific component. Please correct the table or the text and ensure every row reflects the configuration it claims to represent.
  2. [Section 4.3.3, Table 1 vs Section 4.3.5, Table 2] The CA-only configuration appears in both tables with conflicting results. Table 1 reports precision 82.4%, recall 91.8%, mAP_0.5 94.2%, mAP_0.5:0.95 74.8%, parameters 37.2M, and speed 37.5ms. Table 2 reports precision 91.6%, recall 90.5%, mAP_0.5 93.8%, mAP_0.5:0.95 73.4%, parameters 37.7M, and speed 38.6ms. If both rows correspond to 'original model plus CA at the same location', these results should be identical; the discrepancy indicates either different experimental protocols or run-to-run variability. In addition, the CA-only row in Table 1 shows a striking precision drop from 92.0% to 82.4% alongside an mAP_0.5 increase, which is not discussed. This pattern suggests training instability and reinforces the need for multiple seeds.
  3. [Section 4.3, Tables 1 and 3] All reported results are based on a single training run per configuration, with no random seeds, no error bars, and no statistical tests. On a dataset of 3000 images with a 6:2:2 split (600 test images), the central claim of a 1.6-point mAP_0.5 improvement (93.3% to 94.9%) may be within run-to-run variance for a 100-epoch YOLO training. To support the claim that the modifications cause a real improvement, the authors should provide results from multiple independent runs (e.g., at least 3–5 seeds) reported as mean ± standard deviation, and ideally a paired significance test. Without this, the magnitude of the reported gains cannot be distinguished from noise.
  4. [Section 2 and Section 6] The dataset is private and the paper does not release code, trained models, or an evaluation script. Combined with the single-run experimental design, this means the central result is not independently verifiable. The authors should at least provide a detailed evaluation protocol (exact split, hyperparameters, data preprocessing, and inference settings) and, if possible, release anonymized metadata or a small public sample. If the data cannot be shared, the reproducibility argument must rest on rigorous multi-seed reporting and a corrected, internally consistent ablation.
minor comments (6)
  1. [Throughout] The manuscript contains numerous typographical errors and spacing artifacts (e.g., 'insulati on', 'sup ports', 'co exist', 'o bject') that should be corrected in revision.
  2. [Section 4.2] The evaluation metrics are cited to the authors' own prior publication [30] rather than defined in the paper; it would be more helpful to define them briefly or cite the original sources.
  3. [Section 4.3.5, Table 2] The sentence 'the CA attention mechanism achieves the highest accuracy on mAP 0.5 and mAP0.5:0.95 is higher than other attention mechanisms' is grammatically confusing; more importantly, CA's mAP_0.5:0.95 (73.4%) is lower than the original model's 73.5%, so the claim should be qualified to refer only to comparisons among attention mechanisms.
  4. [Section 4.3.7, Table 3] The statement 'The performance of other versions can be estimated proportionally' is speculative and unsupported; please remove it or justify it with evidence.
  5. [Section 5, Conclusion] The conclusion states that the improved model 'accurately detects self-exploded missing and partially damaged insulators', but Table 4 shows recall of only 85.6% for self-explosion defects; 'accurately' overstates the results and should be tempered.
  6. [Section 4.3.4, Figure 12] The attention visualization comparison is purely qualitative. A quantitative measure, such as the overlap between attention maps and ground-truth defect regions, would substantially strengthen the claim that coordinate attention improves focus on small-scale defects.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: RFB, CA, and WIoU are external modules; the reported gains are empirical results, not derivations from the claim itself.

full rationale

The central claim is an experimental comparison on a private dataset. The three modifications are taken from independent prior work (RFB [25], CA [26], WIoU [27]) and are not defined in terms of the paper's mAP or precision/recall results. Equations (1)-(3) state the standard CA formulation and Equations (4)-(8) state the standard WIoUv3 formulation; neither contains the paper's target result as an input. The only self-citation is Section 4.2, where standard metric definitions are said to be found in the authors' earlier ref. [30]; this is not load-bearing because precision, recall, mAP_0.5, and mAP_0.5:0.95 are standard detection metrics whose definitions do not supply or force the reported improvements. The ablation table has an internal consistency problem in the RFB-only row (the text's parameter/FLOP/speed reductions correspond to the RFB+CA row rather than the RFB row), and the evaluation uses a single run without seeds on a non-public dataset; these are reproducibility and reporting concerns, not circularity. No prediction is a renamed fit, no uniqueness theorem is imported, and no equation reduces to its own input, so no circular step is present.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities. Its central claim relies on three domain assumptions about data representativeness, training stability, and baseline fairness, plus several hyperparameters whose values are not disclosed. The WIoU and CA modules bring their own hyperparameters from prior work, and the private dataset prevents external validation.

free parameters (3)
  • WIoUv3 hyperparameters delta and alpha = not stated
    Equations (6)-(8) introduce δ and α to control the outlier degree β and gradient gain r. The paper does not state values used or whether they were tuned; likely inherited from [27].
  • CA reduction ratio r = not stated
    Section 3.3 introduces r as a reduction ratio to control the block size; value not reported.
  • Training hyperparameters = not stated
    Section 4.1 lists hardware and software but not batch size, learning rate, optimizer, or data augmentation choices; training length is fixed at 100 epochs.
assumptions (3)
  • domain assumption The private dataset is representative of real insulator defects and is correctly annotated.
    Section 2: data provided by Yunnan Power Grid with three classes; no external verification or inter-annotator agreement.
  • domain assumption Reported metrics from a single training run are stable enough to compare models.
    Section 4.3 reports single runs without seeds or error bars; the entire comparison depends on this assumption.
  • domain assumption YOLOv7 baseline is a fair and correctly tuned baseline.
    Baseline results in Tables 1-3 are reproduced by the authors, but no external check or hyperparameter details are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improved YOLOv7 model for insulator defect detection." pith.science (2026). https://pith.science/paper/NGBIWS7L

@misc{pith2026250207179,
  author       = {Pith},
  title        = {Pith review of: Improved YOLOv7 model for insulator defect detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NGBIWS7L}},
  note         = {Machine review of arXiv:2502.07179}
}
read the original abstract

Insulators are crucial insulation components and structural supports in power grids, playing a vital role in the transmission lines. Due to temperature fluctuations, internal stress, or damage from hail, insulators are prone to injury. Automatic detection of damaged insulators faces challenges such as diverse types, small defect targets, and complex backgrounds and shapes. Most research for detecting insulator defects has focused on a single defect type or a specific material. However, the insulators in the grid's transmission lines have different colors and materials. Various insulator defects coexist, and the existing methods have difficulty meeting the practical application requirements. Current methods suffer from low detection accuracy and mAP0.5 cannot meet application requirements. This paper proposes an improved YOLOv7 model for multi-type insulator defect detection. First, our model replaces the SPPCSPC module with the RFB module to enhance the network's feature extraction capability. Second, a CA mechanism is introduced into the head part to enhance the network's feature representation ability and to improve detection accuracy. Third, a WIoU loss function is employed to address the low-quality samples hindering model generalization during training, thereby improving the model's overall performance. The experimental results indicate that the proposed model exhibits enhancements across various performance metrics. Specifically, there is a 1.6% advancement in mAP_0.5, a corresponding 1.6% enhancement in mAP_0.5:0.95, a 1.3% elevation in precision, and a 1% increase in recall. Moreover, the model achieves parameter reduction by 3.2 million, leading to a decrease of 2.5 GFLOPS in computational cost. Notably, there is also an improvement of 2.81 milliseconds in single-image detection speed.

Figures

Figures reproduced from arXiv: 2502.07179 by the authors.

Figure 6
Figure 6. Comparison of mAP_0.5 [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Comparison of mAP_0.5:0.95 [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 9
Figure 9. Comparison of recall. 4.3.2. Detection result [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figures from the paper (4 more)
Figure 10
Figure 10. Figure 10: Detection results of the original YOLOv7 model [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Detection results of our improved YOLOv7 model. Because the detected results have large sizes with a resolution of 4000 * 3000 pixels and the detected insulator is only a tiny proportion of the whole image, it isn't easy to see the actual label after detection. In [P…
Figure 12
Figure 12. Figure 12: Comparison of attentional visualization. Based on the comparison between [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Comparison of IoU loss function. As shown in [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 25 canonical work pages

  1. [1]

    Peng, F.X

    X.Y . Peng, F.X. Liang, J.J. Qian, B.S. Yang, C. Chen, X.G. Zheng, Automatic localization of transmission line insulators bas ed on airborne infrared image t exture features, High Voltage Technology, 45 (2019), 922–928. https://doi.org/10.13336/j.1003-6520.hve.2019 0226033 (in Chinese)

    work page doi:10.13336/j.1003-6520.hve.2019 2019

  2. [2]

    Liu, X.R

    Z.Y . Liu, X.R. Liao, J. Chen, H. Jiang, Review on intelligent processing of visible light images for inspection of power overhead line, Power System Technology , 44 (2020), 1057–1069. https://doi.org/10.13335/j.1000-3673.pst.2019.0349 (in Chinese)

    work page arXiv 2020

  3. [3]

    Huang, Z.S

    S. Huang, Z.S. Wu, Z.G. Ren, H.J. Liu, Y . Gui, Review of resear ch on intelligent inspection robots for electric power, Electrical Measurement and Instrumentation , 57 (2020), 26–38, https://doi.org/10.19753/j.issn1001-1390.2020.002.005 (in Chinese)

    work page arXiv 2020

  4. [4]

    Liu, L.H

    G. Liu, L.H. Wu, H. Zhang, Research on key technologies for int elligent inspection of helicopters on transmission lines, Journal of Three Gorges Un iversity (Natural Science) , 36 (2014), 46-49+62. https://doi.org/10.13393/j.cnki.issn.1672-948x.2014.02.011 (in Chinese)

    work page doi:10.13393/j.cnki.issn.1672-948x.2014.02.011 2014

  5. [5]

    Chen, Y .H

    F.M. Chen, Y .H. Du, H. Chen, Application of image processing te chnology in intelligent inspection of helicopters on transmission lines, Zhejiang Electric Power , 31 (2012), 63-66. https://doi.org/10.19585/j.zjdl.2012.09.018 (in Chinese) 18

    work page doi:10.19585/j.zjdl.2012.09.018 2012

  6. [6]

    Yan, Measurement model for ice thickness of transmission lines based on image technology, Electrical Engineering Materials , (2021), 66-69+72

    H. Yan, Measurement model for ice thickness of transmission lines based on image technology, Electrical Engineering Materials , (2021), 66-69+72. https://doi.org/10.16786/j.cnki.1671-8887.eem.2021.05.018 (in Chinese)

    work page doi:10.16786/j.cnki.1671-8887.eem.2021.05.018 2021

  7. [7]

    Yan, F.C

    K. Yan, F.C. Wang, C.Y . Zhang, Edge detection of composite insulator hydrophobicity images based on Canny operator, Journal of Electric Po wer Science and Technology , 28 (2013), 45– 49+56. https://doi.org/10.3969/j.issn.1673-9140.2013.03.006 (in Chinese)

    work page arXiv 2013

  8. [8]

    Felzenszwalb, R.B

    P.F. Felzenszwalb, R.B. Girshick, D. McAllester, D. Ramanan, Ob ject Detection with Discriminatively Trained Part-Based Models, IEEE Transactions on Pa ttern Analysis and Machine Intelligence, 32 (2010), 1627–1645, https://doi.org/10.1109/TPAMI.2009.167

    doi:10.1109/tpami.2009.167 2010

Show all 31 references
  1. [9]

    Liu, B.Q

    K.P. Liu, B.Q. Li, L. Qin, Q. Li, F. Zhao, Q.L. Wang, et.al., R eview on the application of deep learning target detection algorit hm in insulator defect detecti on of overhead transmission line, High Voltage Technology , 49 (2023), 3584–3595. https://doi.org/10.13336/j.1003-6...

    2023 doi

  2. [10]

    Jiang, X

    H. Jiang, X. Qiu, J. Chen, X. Liu, X. Miao, S. Zhuang, Insulato r fault detection in aerial images based on ensemble learning w ith multi-level perception, IEEE Access, 7 (2019), 61797–61810. https://doi.org/10.1109/ACCESS.2019.2915985

    2019

  3. [11]

    Sadykova, D

    D. Sadykova, D. Pernebayeva, M. B agheri, A. James, IN-YOLO: Rea l-time detection of outdoor high voltage insulators using UA V imaging, IEEE Transactions on Power Delivery, 35 (2020), 1599–1601. https://doi.org/10.1109/TPWRD.2019.2944741

    2020

  4. [12]

    Wang, Y .F

    S.Q. Wang, Y .F. Liu, Y .H. Qing, C.X. Wang, T.Z. Lan, R.T. Yao, Detection of insulator defects with improved ResNeSt and region proposal network, IEEE Access, 8 (2020), 184841–184850. https://doi.org/10.1109/ACCESS.2020.3029857

    2020

  5. [13]

    Shi, Y .P

    C.X. Shi, Y .P. Huang, Cap-count guided weakly supervised insulator cap missing detection in aerial images, IEEE Sensors Journal , 21 (2021), 685–691. https://doi.org/10.1109/JSEN.2020.3012780

    2021

  6. [14]

    Zhao, M.F

    W.Q. Zhao, M.F. Xu, X.F. Cheng, Z.B. Zhao, An insulator in transmission lines recognition and fault detection model based on improved faster rcnn, IEEE Transactions on Instrumentation and Measurement, 70 (2021), 1–8. https://doi.org/10.1109/TIM.2021.3112227

    2021

  7. [15]

    P . Luo, B. W ang, H.R. Ma, F .Q. Ma, H.X. W ang, D.H. Zhu, Low miss rate defect identification method based on combined target detection framework, High Voltage Technology, 47 (2021), 454–464. https://doi.org/10.13336/j.1003-6520.hve.20200701 (in Chinese)

    2021 doi

  8. [16]

    Zhang, Y .Y

    X.T. Zhang, Y .Y . Zhang, J.F. Liu, C.H. Zhang, X.Y . Xue, H. Zhang, et al., InsuDet: A fault detection method for insulators of overhead transmission lines using convolutional neural networks, IEEE Transactions on Instrumentation and Measurement , 70 (2021), 1–12. https://doi...

    2021

  9. [17]

    Liu, X.B

    Y . Liu, X.B. Huang, Research on insulator burst detection and localization based on YOLOv4 and improved watershed algorithm, Power System and Clean Energy , 37 (2021), 51–57. https://doi.org/10.3969/j.issn.1674-3814.2021.07.007 (in Chinese)

    2021

  10. [18]

    Liu, X.R

    X.Y . Liu, X.R. Liao, S.B. Zhuang, H. Jiang, J. Chen, Insulator detection based on lightweight deep convolutional neural network, Journal of Fuzhou University (Natural Science Edition), 49 (2021), 196–202. https://doi.org/10.7631/issn.1000-2243.20345 (in Chinese) 19

    2021

  11. [19]

    Wang, J.J

    D.L. Wang, J.J. Sun, T.Y . Zhuang, M.S. Li, R. Zhu, Detection method for self-explosion defects of glass insulators based on improved generative adversarial networks, High Voltage Technology, 48 (2022), 1096–1103. https://doi.org/10.13336/j.1003-6520.hve.20210236 (in Chinese)

    2022 doi

  12. [20]

    J. Kang, Q. Wang, W.B. Liu, Y . Xia, Multi defect detection netw ork for aerial insulators integrating CAT-BiFPN and attention mechanism, High Voltage Technology, 49 (2023), 3361–

    2023

  13. [21]

    Singh, S.F

    G. Singh, S.F. Stefenon, K.C. Yow, Interpretable visual transmi ssion lines inspections using pseudo-prototypical part network, Machine Vision and Applications , 41 (2023). https://doi.org/10.1007/s00138-023-01390-6

    2023 doi

  14. [22]

    Souza, S.F

    B.J. Souza, S.F. Serfenon, G. Singh, R.Z. Freire, Hybrid-YOLO f or classification of insulators defects in transmission lines based on UA V , International Journal of Electrical Power and Energy Systems, 148 (2023), 108982. https://doi.org/10.1016/j.ijepes.2023.108982

    2023

  15. [23]

    Serfenon, G

    S.F. Serfenon, G. Singh, B.J. Souza, R.Z. Freire, K.C. Yow, Opt imized hybrid YOLOu-Quasi-ProtoPNet for i nsulators classification, IET Generation, Transmission and Distribution, 17 (2023), 3501–3511. https://doi.org/10.1049/gtd2.12886

    2023 doi

  16. [24]

    C.Y . Wang, A. Bochkovskiy, H.Y .M. Liao, YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real- time object detectors, in 2022 IEEE/CVF Conference on Computer Vision and Pattern R ecognition (CVPR) , (2022), 7464–7475. https://doi.org/10.48550/arXiv.2207.02696

    2022 doi arXiv

  17. [25]

    S. Liu, D. Huang, Y .H. Wang, Receptive field block net for accurate and fast object detection, in 2018 European Conference on Co mputer Vision (ECCV) , (2018), 385–400. https://doi.org/10.1007/978-3-030-01252-6_24

    2018 doi

  18. [26]

    Hou, D.Q

    Q.B. Hou, D.Q. Zhou, J.S. Feng, C oordinate attention for effici ent mobile network design, in 2021 EEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR) , (2021), 13708–13717. https://doi.org/10.48550/arXiv.2103.02907

    2021 doi arXiv

  19. [27]

    Tong, Y .H

    Z.J. Tong, Y .H. Chen, Z.W. Xu, R. YU, Wise-IoU: Bounding box regression loss with dynamic focusing mechanism, in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), (2023). https://doi.org/10.48550/arXiv.2301.10051

    2023 doi arXiv

  20. [28]

    Szegedy, W

    C. Szegedy, W. Liu, Y .Q. Jia, P. Sermanet, S. Reed, D. Anguelov, et al., Going deeper with convolution, in 2015 EEE/CVF Conference on Comput er Vision and Pattern Recognition (CVPR), (2015), 1–9. https://doi.org/10.1109/CVPR.2015.7298594

    2015

  21. [29]

    F. Yu, V . Koltun, Multi-Scale context aggregation by dilated convolutions, in 2016 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , (2015). https://doi.org/10.48550/arXiv.1511.07122

    2015 doi arXiv

  22. [30]

    Wang, G.W

    Z.Y . Wang, G.W. Yuan, H. Zhou, Y . Ma, Y .T. Ma. Foreign-Object Detection in High-V oltage Transmission Line Based on Improved YOLOv8m, Applied Sciences , 13(2023), 12775. https://doi.org/10.3390/app132312775

    2023 doi

  23. [3376]

    https://doi.org/10.13336/j.1003-6520.hve.20221803 (in Chinese)

    doi

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.