pith. sign in

arxiv: 2606.12958 · v1 · pith:GDQDKWZVnew · submitted 2026-06-11 · 💻 cs.CV

YOLO-AMC: An Improved YOLO Architecture with Attention Mechanisms for Building Crack Detection

Ching-Yu Tsai , Chia-Min Lin , Chih-Hsiang Yang , Yung-Che Wang , Jen-Shiun Chiang This is my paper

Pith reviewed 2026-06-27 07:32 UTC · model grok-4.3

classification 💻 cs.CV
keywords crack detectionYOLOattention mechanismsobject detectionstructural health monitoringbuilding inspectionGAMcomputer vision
0
0 comments X

The pith

Inserting the Global Attention Mechanism into YOLOv11's neck after C2PSA removal raises crack detection mAP@0.5 to 0.9917.

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

The paper modifies YOLOv11 by removing the C2PSA module and placing attention mechanisms into the neck's multi-scale feature fusion layers to better handle thin, low-contrast cracks amid background noise. It tests GAM, Res-CBAM, and SA, with GAM delivering the highest scores. These changes aim to improve automated detection for infrastructure inspection without increasing computational cost much. A reader would care if the gains translate to more reliable real-world monitoring of building damage.

Core claim

YOLO-AMC with GAM inserted into the multi-scale feature fusion layers of the Neck after C2PSA removal achieves mAP@0.5 of 0.9917 and mAP@0.5:0.95 of 0.9506 on the test dataset, exceeding YOLOv11's 0.9833 and 0.9112 as well as YOLOv8's 0.9707 and 0.8921, while operating at 7.6 GFLOPs, 110.95 FPS on RTX 4090, and roughly 5 FPS on Raspberry Pi 5.

What carries the argument

Attention modules placed in the multi-scale feature fusion layers of the Neck to strengthen cross-scale feature integration for low-contrast crack detection.

If this is right

  • GAM outperforms the other tested attention modules for this crack detection task.
  • The resulting model maintains real-time speeds on both desktop GPUs and edge hardware such as the Raspberry Pi 5.
  • The architecture offers a direct accuracy improvement over standard YOLOv11 and YOLOv8 for thin-structure detection in noisy images.
  • Deployment remains feasible at low computational cost of 7.6 GFLOPs.

Where Pith is reading between the lines

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

  • The same neck-level attention placement could be tested on detection of other low-contrast linear features such as roads or pipelines in aerial imagery.
  • If the gain depends on cross-scale fusion, applying the modules earlier in the backbone might yield different trade-offs worth measuring.
  • The approach could be combined with data augmentation targeted at crack-like artifacts to see whether the mAP ceiling rises further.

Load-bearing premise

The reported accuracy gains arise specifically from the attention module insertions and C2PSA removal rather than from unstated choices in training data, optimization, or other implementation details.

What would settle it

Reproduce the exact training protocol on the same dataset using unmodified YOLOv11 without the attention additions or C2PSA removal, then measure whether its mAP@0.5 and mAP@0.5:0.95 match the lower baseline values or rise to the reported YOLO-AMC levels.

Figures

Figures reproduced from arXiv: 2606.12958 by Chia-Min Lin, Chih-Hsiang Yang, Ching-Yu Tsai, Jen-Shiun Chiang, Yung-Che Wang.

Figure 1
Figure 1. Figure 1: Overall architecture of the proposed YOLO-AMC framework and the attention insertion positions in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the proposed YOLO-AMC model. Attention modules are inserted at different positions [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of detection results on a crack image. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of detection results on a crack image. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of feature response heatmaps at different feature levels (P3, P4, P5). [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of crack detection results of the YOLO-AMC (GAM) model on GPU and Raspberry Pi 5 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
read the original abstract

Crack detection plays an important role in infrastructure inspection and Structural Health Monitoring (SHM). However, cracks typically appear as thin, low-contrast structures and are easily affected by background noise, posing challenges for existing object detection models. This study proposes an improved YOLO-based architecture with integrated attention mechanisms, termed YOLO-AMC (YOLO with Attention Mechanisms for Crack Detection), to enhance automated crack detection performance. Based on YOLOv11, the original C2PSA module is removed, and multiple attention mechanisms, including Global Attention Mechanism (GAM), Residual Convolutional Block Attention Module (Res-CBAM), and Shuffle Attention (SA), are introduced into the multi-scale feature fusion layers of the Neck to strengthen cross-scale feature integration. Experimental results demonstrate that YOLO-AMC consistently outperforms baseline models YOLOv11n and YOLOv8n across multiple evaluation metrics. Among the evaluated attention modules, GAM achieves the best detection performance, obtaining mAP@0.5 = 0.9917 and mAP@0.5:0.95 = 0.9506 on the test dataset, which are higher than those of YOLOv11 (0.9833 / 0.9112) and YOLOv8 (0.9707 / 0.8921). Furthermore, while maintaining a computational complexity of 7.6 GFLOPs, the proposed model achieves 110.95 FPS on an NVIDIA RTX 4090 platform and approximately 5 FPS on a Raspberry Pi 5 edge device, demonstrating a favorable trade-off between accuracy and deployment efficiency. The implementation code for this study is available on GitHub at https://github.com/CY-Tsai24/YOLO-AMC.

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

1 major / 2 minor

Summary. The paper proposes YOLO-AMC, an improved YOLOv11-based detector for building crack detection. It removes the original C2PSA module and inserts attention mechanisms (GAM, Res-CBAM, or SA) into the multi-scale feature fusion layers of the Neck. On a held-out test set the GAM variant reports mAP@0.5 = 0.9917 and mAP@0.5:0.95 = 0.9506, exceeding the unmodified YOLOv11n (0.9833 / 0.9112) and YOLOv8n (0.9707 / 0.8921) baselines, while retaining 7.6 GFLOPs and 110.95 FPS on an RTX 4090 (≈5 FPS on Raspberry Pi 5). Public code is provided.

Significance. If the reported gains can be isolated to the attention insertions, the work would supply a concrete, deployable improvement for low-contrast crack detection in structural health monitoring. The public GitHub link and edge-device FPS numbers are concrete strengths that would aid reproducibility and practical adoption.

major comments (1)
  1. [Abstract and Experimental Results] Abstract and Experimental Results: the central claim attributes the mAP@0.5 / mAP@0.5:0.95 gains specifically to insertion of GAM (or Res-CBAM/SA) after C2PSA removal. However, the only reported comparisons are the full YOLO-AMC model versus unmodified YOLOv11n and YOLOv8n; no ablation table or section shows (a) YOLOv11 with C2PSA removed but no attention added, (b) attention added without C2PSA removal, or (c) identical training schedule, augmentations, and optimizer across variants. Without these controls the performance delta cannot be confidently assigned to the attention modules.
minor comments (2)
  1. The manuscript does not state the number of images in the training/validation/test splits or the precise training protocol (optimizer, learning-rate schedule, augmentation policy, number of epochs).
  2. Table or figure captions should explicitly list the exact attention-module variants evaluated and the baseline configurations.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. The key concern is the absence of ablation studies needed to attribute performance gains specifically to the attention insertions after C2PSA removal. We address this point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract and Experimental Results] Abstract and Experimental Results: the central claim attributes the mAP@0.5 / mAP@0.5:0.95 gains specifically to insertion of GAM (or Res-CBAM/SA) after C2PSA removal. However, the only reported comparisons are the full YOLO-AMC model versus unmodified YOLOv11n and YOLOv8n; no ablation table or section shows (a) YOLOv11 with C2PSA removed but no attention added, (b) attention added without C2PSA removal, or (c) identical training schedule, augmentations, and optimizer across variants. Without these controls the performance delta cannot be confidently assigned to the attention modules.

    Authors: We agree that the current experiments compare only the complete YOLO-AMC model against the unmodified YOLOv11n and YOLOv8n baselines, without the requested intermediate variants. This limits the ability to isolate the effect of the attention modules. In the revised manuscript we will add a new ablation subsection and table that reports: (a) YOLOv11 with C2PSA removed but no attention modules, (b) YOLOv11 with the attention modules inserted while retaining the original C2PSA, and (c) explicit confirmation that all variants (including baselines) were trained with identical schedules, augmentations, optimizer, learning-rate policy, and epoch count. These additional results will be generated from the publicly released code and will be used to strengthen the attribution of gains to the proposed modifications. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical mAP results on held-out test set with no derivations or self-referential fits

full rationale

The paper proposes an architectural change (C2PSA removal + attention insertion in Neck) and reports measured mAP values on a test dataset. No equations, first-principles derivations, or predictions appear that reduce the reported metrics to quantities defined by the paper's own fitted parameters or self-citations. The central claim is an empirical comparison against unmodified baselines; absence of ablations affects evidential strength but does not create circularity by construction. No self-citation load-bearing steps, ansatz smuggling, or renaming of known results are present in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Minimal ledger because the work is an empirical architecture tweak rather than a theoretical derivation; the sole domain assumption is that attention modules will improve feature focus on thin low-contrast structures.

axioms (1)
  • domain assumption Attention mechanisms improve cross-scale feature integration for thin, low-contrast objects such as cracks.
    Invoked to justify placement of GAM, Res-CBAM, and SA in the Neck.

pith-pipeline@v0.9.1-grok · 5867 in / 1175 out tokens · 18826 ms · 2026-06-27T07:32:02.108619+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

48 extracted references · 5 canonical work pages · 4 internal anchors

  1. [1]

    Y.-J. Cha, W. Choi, G. Suh, S. Mahmoudkhani, and O. B¨ uy¨ uk¨ ozt¨ urk. Autonomous structural vi- sual inspection using region-based deep learning for detecting multiple damage types.Computer-Aided Civil and Infrastructure Engineering, 33(9):731–747, 2018

    2018

  2. [2]

    R. Fan, M. J. Bocus, Y. Zhu, J. Jiao, L. Wang, F. Ma, S. Cheng, and M. Liu. Road crack detec- tion using deep convolutional neural network and adaptive thresholding. InProc. IEEE Intelligent Vehicles Symposium (IV), pages 474–479, 2019

    2019

  3. [3]

    Smart structural health monitoring using computer vision and edge computing.Engineering Structures, 319:118809, 2024

    Zhen Peng, Jun Li, Hong Hao, and Yue Zhong. Smart structural health monitoring using computer vision and edge computing.Engineering Structures, 319:118809, 2024

    2024

  4. [4]

    S. Y. Mohammed. Architecture review: Two-stage and one-stage object detection.Franklin Open, page 100322, 2025

    2025

  5. [5]

    M. L. Ali and Z. Zhang. The yolo framework: A comprehensive review of evolution, applications, and benchmarks in object detection.Computers, 13(12):336, 2024

    2024

  6. [6]

    Ultralytics yolo (version 8.0.0)

    Ultralytics. Ultralytics yolo (version 8.0.0). https: //github.com/ultralytics/ultralytics, 2025. Accessed: Jan. 13, 2025

    2025

  7. [7]

    A lightweight underwater detector enhanced by attention mechanism, gsconv and wiou on yolov8

    Shaobin Cai, Xiangkui Zhang, and Yuchang Mo. A lightweight underwater detector enhanced by attention mechanism, gsconv and wiou on yolov8. Scientific Reports, 14(1):25797, 2024

    2024

  8. [8]

    Alzahrani, Arif Bramantoro, and Mira Kartiwi

    Fatin Najihah Muhamad Zamri, Teddy Surya Gu- nawan, Siti Hajar Yusoff, Ahmad A. Alzahrani, Arif Bramantoro, and Mira Kartiwi. Enhanced small 12 drone detection using optimized yolov8 with atten- tion mechanisms.IEEE Access, 12:90629–90643, 2024

    2024

  9. [9]

    N. S. Rupak, N. Rayvanth, Pulipati Kushank, and Rimjhim Padam Singh. An investigation into yolo- v8 model optimization for small object detection in uavs using attention mechanism. In2024 15th International Conference on Computing Commu- nication and Networking Technologies (ICCCNT), pages 1–7. IEEE, 2024

    2024

  10. [10]

    Lai, C.-Y

    J.-Y. Lai, C.-Y. Tsai, C. H. Yang, C.-M. Lin, and J.-S. Chiang. Yolo with attention mechanisms for building cracks detection. InProc. IET Interna- tional Conference on Engineering Technologies and Applications (ICETA), 2025

    2025

  11. [11]

    Y.-J. Cha, W. Choi, and O. B¨ uy¨ uk¨ ozt¨ urk. Deep learning-based crack damage detection using convo- lutional neural networks.Computer-Aided Civil and Infrastructure Engineering, 32(5):361–378, 2017

    2017

  12. [12]

    Z. Fan, Y. Wu, J. Lu, and W. Li. Automatic pavement crack detection based on structured prediction with the convolutional neural network. arXiv:1802.02208, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  13. [13]

    S. Li, X. Zhao, and G. Zhou. Automatic pixel-level multiple damage detection of concrete structure using fully convolutional network.Computer-Aided Civil and Infrastructure Engineering, 34(7):616–634, 2019

    2019

  14. [14]

    F. Yang, L. Zhang, S. Yu, D. Prokhorov, X. Mei, and H. Ling. Feature pyramid and hierarchical boosting network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 21(4):1525–1535, 2019

    2019

  15. [15]

    H. Li, P. Xiong, J. An, and L. Wang. Pyra- mid attention network for semantic segmentation. arXiv:1805.10180, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  16. [16]

    YOLOv3: An Incremental Improvement

    J. Redmon and A. Farhadi. Yolov3: An incremental improvement. arXiv:1804.02767, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  17. [17]

    YOLOv4: Optimal Speed and Accuracy of Object Detection

    A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao. Yolov4: Optimal speed and accuracy of object de- tection. arXiv:2004.10934, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2004

  18. [18]

    Z. Zhou, J. Zhang, C. Gong, and W. Wu. Automatic tunnel lining crack detection via deep learning with generative adversarial network-based data augmen- tation.Underground Space, 9:140–154, 2023

    2023

  19. [19]

    X. Li, N. Zhang, Y. Pan, Y. Lv, X. Xu, and Z. Wang. A lightweight and attention-enhanced framework for robust pavement defect detection.Engineering Applications of Artificial Intelligence, 165:113545, 2026

    2026

  20. [20]

    Liang, H

    S. Liang, H. Wu, L. Zhen, Q. Hua, S. Garg, G. Kad- doum, M. M. Hassan, and K. Yu. Edge yolo: Real-time intelligent object detection system based on edge-cloud cooperation in autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 23(12):25345–25360, 2022

    2022

  21. [21]

    S. Woo, J. Park, J.-Y. Lee, and I. S. Kweon. Cbam: Convolutional block attention module. InProceed- ings of the European Conference on Computer Vi- sion (ECCV), pages 3–19, 2018

    2018

  22. [22]

    Y. Liu, Z. Shao, and N. Hoffmann. Global attention mechanism: Retain information to enhance channel- spatial interactions. arXiv:2112.05561, 2021

    work page arXiv 2021

  23. [23]

    Zhang and Y.-B

    Q.-L. Zhang and Y.-B. Yang. Sa-net: Shuffle atten- tion for deep convolutional neural networks. InPro- ceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2235–2239, 2021

    2021

  24. [24]

    F. Wang, M. Jiang, C. Qian, S. Yang, C. Li, H. Zhang, X. Wang, and X. Tang. Residual at- tention network for image classification. InProceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3156–3164, 2017

    2017

  25. [25]

    J. Hu, L. Shen, and G. Sun. Squeeze-and-excitation networks. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7132–7141, 2018

    2018

  26. [26]

    Chien, R.-Y

    C.-T. Chien, R.-Y. Ju, K.-Y. Chou, E. Xieerke, and J.-S. Chiang. Yolov8-am: Yolov8 based on effective attention mechanisms for pediatric wrist fracture detection.IEEE Access, 2025

    2025

  27. [27]

    X. Dong, J. Yuan, and J. Dai. Study on lightweight bridge crack detection algorithm based on yolo11. Sensors, 25(11):3276, 2025

    2025

  28. [28]

    M. Kim, J. Jeong, and S. Kim. Ecap-yolo: Ef- ficient channel attention pyramid yolo for small object detection in aerial image.Remote Sensing, 13(23):4851, 2021

    2021

  29. [29]

    S. Tang, S. Zhang, and Y. Fang. Hic-yolov5: Im- proved yolov5 for small object detection. InPro- ceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 6614–6619, 2024

    2024

  30. [30]

    G. Wang, H. Ding, Z. Yang, B. Li, Y. Wang, and L. Bao. Trc-yolo: A real-time detection method for lightweight targets based on mobile devices.IET Computer Vision, 16(2):126–142, 2022

    2022

  31. [31]

    T. Saeheaw. Hfe-yolo: Hybrid feature enhancement with multi-attention mechanisms for construction site object detection.Buildings, 15(23):4274, 2025

    2025

  32. [32]

    X. Li, W. Wang, X. Hu, and J. Yang. Selective kernel networks. InProceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 510–519, 2019. 13

    2019

  33. [33]

    T.-Y. Lin, P. Doll´ ar, R. Girshick, K. He, B. Hari- haran, and S. Belongie. Feature pyramid networks for object detection. InProceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), pages 2117–2125, 2017

    2017

  34. [34]

    M. Tan, R. Pang, and Q. V. Le. Efficientdet: Scal- able and efficient object detection. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10781–10790, 2020

    2020

  35. [35]

    Jiang, C

    T. Jiang, C. Li, M. Yang, and Z. Wang. An im- proved yolov5s algorithm for object detection with an attention mechanism.Electronics, 11(16):2494, 2022

    2022

  36. [36]

    W. Li, K. Liu, L. Zhang, and F. Cheng. Object de- tection based on an adaptive attention mechanism. Scientific Reports, 10(1):11307, 2020

    2020

  37. [37]

    Zhang, Y

    Y. Zhang, Y. Chen, C. Huang, and M. Gao. Object detection network based on feature fusion and at- tention mechanism.Future Internet, 11(1):9, 2019

    2019

  38. [38]

    Bac hien crack concrete 2024 dataset

    Roboflow. Bac hien crack concrete 2024 dataset. Roboflow Universe, 2024. [Accessed: Feb. 10, 2025]

    2024

  39. [39]

    Crack detection.v2 dataset

    Roboflow. Crack detection.v2 dataset. Roboflow Universe, 2024. [Accessed: Feb. 10, 2025]

    2024

  40. [40]

    Crack detection.v3i dataset

    Roboflow. Crack detection.v3i dataset. Roboflow Universe, 2024. [Accessed: Feb. 10, 2025]

    2024

  41. [41]

    Crack finder.v1i dataset

    Roboflow. Crack finder.v1i dataset. Roboflow Uni- verse, 2024. [Accessed: Feb. 10, 2025]

    2024

  42. [42]

    C. F. ¨Ozgenel and A. Sorgu¸ c. Concrete crack images for classification dataset. Mendeley Data / Kaggle,

  43. [43]

    3, 2025]

    [Accessed: Feb. 3, 2025]

    2025

  44. [44]

    Chattopadhay, A

    A. Chattopadhay, A. Sarkar, P. Howlader, and V. N. Balasubramanian. Grad-cam++: General- ized gradient-based visual explanations for deep convolutional networks. InProceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839–847. IEEE, 2018

    2018

  45. [45]

    Zeiler and Rob Fergus

    Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. InEuropean Conference on Computer Vision (ECCV), pages 818–833. Springer, 2014

    2014

  46. [46]

    Wichmann, and Wieland Brendel

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias im- proves accuracy and robustness. InInternational Conference on Learning Representations (ICLR), 2019

    2019

  47. [47]

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discrim- inative localization. InProceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), pages 2921–2929, 2016

    2016

  48. [48]

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra. Grad-cam: Visual ex- planations from deep networks via gradient-based localization. InProceedings of the IEEE Inter- national Conference on Computer Vision (ICCV), pages 618–626, 2017. 14

    2017