Model generalization experiment.
<div><p>Road crack detection is critical to global infrastructure maintenance and public safety, and complex background environments and nonlinear damage crack patterns challenge the need for real-time, efficient, and accurate detection.This paper proposes a lightweight yet robust Flexi-...
محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| مؤلفون آخرون: | , , , |
| منشور في: |
2025
|
| الموضوعات: | |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
| _version_ | 1852019294599118848 |
|---|---|
| author | Jiexiang Yang (19980594) |
| author2 | Renjie Tian (21553556) Zexing Zhou (21553559) Xingyue Tan (21553562) Pingyang He (21553565) |
| author2_role | author author author author |
| author_facet | Jiexiang Yang (19980594) Renjie Tian (21553556) Zexing Zhou (21553559) Xingyue Tan (21553562) Pingyang He (21553565) |
| author_role | author |
| dc.creator.none.fl_str_mv | Jiexiang Yang (19980594) Renjie Tian (21553556) Zexing Zhou (21553559) Xingyue Tan (21553562) Pingyang He (21553565) |
| dc.date.none.fl_str_mv | 2025-06-16T17:37:21Z |
| dc.identifier.none.fl_str_mv | 10.1371/journal.pone.0325993.t006 |
| dc.relation.none.fl_str_mv | https://figshare.com/articles/dataset/Model_generalization_experiment_/29331674 |
| dc.rights.none.fl_str_mv | CC BY 4.0 info:eu-repo/semantics/openAccess |
| dc.subject.none.fl_str_mv | Ecology Space Science Biological Sciences not elsewhere classified Information Systems not elsewhere classified recall rate rise model &# 8217 lower application costs experimental results show complex shape characteristics complex background environments akconv convolution operation global infrastructure maintenance f1 score improvement lightweight network design road crack detection global attention module yolo model based 5 &# 8211 global information c2f module lightweight method automated detection accurate detection yolov8 algorithm yolo achieves regression accuracy quality samples public safety precise solution paper proposes map improvement loss function industrial demands feature redundancy feature information establishing g designed wise cracks dynamically convolutional kernels bounding boxes adaptively adjust accuracy increase 95 increase 9 %, |
| dc.title.none.fl_str_mv | Model generalization experiment. |
| dc.type.none.fl_str_mv | Dataset info:eu-repo/semantics/publishedVersion dataset |
| description | <div><p>Road crack detection is critical to global infrastructure maintenance and public safety, and complex background environments and nonlinear damage crack patterns challenge the need for real-time, efficient, and accurate detection.This paper proposes a lightweight yet robust Flexi-YOLO model based on the YOLOv8 algorithm. We designed Wise-IoU as the model’s loss function to optimize the regression accuracy of its bounding boxes and enhance robustness to low-quality samples. The DCNv-C2f module is constructed for the transformation and fusion of feature information, allowing the convolutional kernels to adapt to the complex shape characteristics of cracks dynamically. A Global Attention Module (GAM) is integrated to improve the model’s perception of global information. The AKConv convolution operation is employed to adaptively adjust the size of convolutions, further enhancing local feature capturing. Additionally, a lightweight network design is implemented, establishing G-Head (Ghost-Head) as the detection head to optimize the issue of feature redundancy. Experimental results show that Flexi-YOLO achieves an accuracy increase of 2.7% over YOLOv8n, a recall rate rise of 4.7%, a mAP improvement of 5.3%, a mAP@0.5–0.95 increase of 3.9%, a decrease of 0.5 in GFLOPS, and an F1 score improvement from 0.80 to 0.84. Flexi-YOLO offers higher detection accuracy and robustness and meets the industrial demands for lightweight real-time detection and lower application costs, providing an efficient and precise solution for the automated detection of road cracks.</p></div> |
| eu_rights_str_mv | openAccess |
| id | Manara_f04e2e7a9f73403bcb39b39cd325df79 |
| identifier_str_mv | 10.1371/journal.pone.0325993.t006 |
| network_acronym_str | Manara |
| network_name_str | ManaraRepo |
| oai_identifier_str | oai:figshare.com:article/29331674 |
| publishDate | 2025 |
| repository.mail.fl_str_mv | |
| repository.name.fl_str_mv | |
| repository_id_str | |
| rights_invalid_str_mv | CC BY 4.0 |
| spelling | Model generalization experiment.Jiexiang Yang (19980594)Renjie Tian (21553556)Zexing Zhou (21553559)Xingyue Tan (21553562)Pingyang He (21553565)EcologySpace ScienceBiological Sciences not elsewhere classifiedInformation Systems not elsewhere classifiedrecall rate risemodel &# 8217lower application costsexperimental results showcomplex shape characteristicscomplex background environmentsakconv convolution operationglobal infrastructure maintenancef1 score improvementlightweight network designroad crack detectionglobal attention moduleyolo model based5 &# 8211global informationc2f modulelightweight methodautomated detectionaccurate detectionyolov8 algorithmyolo achievesregression accuracyquality samplespublic safetyprecise solutionpaper proposesmap improvementloss functionindustrial demandsfeature redundancyfeature informationestablishing gdesigned wisecracks dynamicallyconvolutional kernelsbounding boxesadaptively adjustaccuracy increase95 increase9 %,<div><p>Road crack detection is critical to global infrastructure maintenance and public safety, and complex background environments and nonlinear damage crack patterns challenge the need for real-time, efficient, and accurate detection.This paper proposes a lightweight yet robust Flexi-YOLO model based on the YOLOv8 algorithm. We designed Wise-IoU as the model’s loss function to optimize the regression accuracy of its bounding boxes and enhance robustness to low-quality samples. The DCNv-C2f module is constructed for the transformation and fusion of feature information, allowing the convolutional kernels to adapt to the complex shape characteristics of cracks dynamically. A Global Attention Module (GAM) is integrated to improve the model’s perception of global information. The AKConv convolution operation is employed to adaptively adjust the size of convolutions, further enhancing local feature capturing. Additionally, a lightweight network design is implemented, establishing G-Head (Ghost-Head) as the detection head to optimize the issue of feature redundancy. Experimental results show that Flexi-YOLO achieves an accuracy increase of 2.7% over YOLOv8n, a recall rate rise of 4.7%, a mAP improvement of 5.3%, a mAP@0.5–0.95 increase of 3.9%, a decrease of 0.5 in GFLOPS, and an F1 score improvement from 0.80 to 0.84. Flexi-YOLO offers higher detection accuracy and robustness and meets the industrial demands for lightweight real-time detection and lower application costs, providing an efficient and precise solution for the automated detection of road cracks.</p></div>2025-06-16T17:37:21ZDatasetinfo:eu-repo/semantics/publishedVersiondataset10.1371/journal.pone.0325993.t006https://figshare.com/articles/dataset/Model_generalization_experiment_/29331674CC BY 4.0info:eu-repo/semantics/openAccessoai:figshare.com:article/293316742025-06-16T17:37:21Z |
| spellingShingle | Model generalization experiment. Jiexiang Yang (19980594) Ecology Space Science Biological Sciences not elsewhere classified Information Systems not elsewhere classified recall rate rise model &# 8217 lower application costs experimental results show complex shape characteristics complex background environments akconv convolution operation global infrastructure maintenance f1 score improvement lightweight network design road crack detection global attention module yolo model based 5 &# 8211 global information c2f module lightweight method automated detection accurate detection yolov8 algorithm yolo achieves regression accuracy quality samples public safety precise solution paper proposes map improvement loss function industrial demands feature redundancy feature information establishing g designed wise cracks dynamically convolutional kernels bounding boxes adaptively adjust accuracy increase 95 increase 9 %, |
| status_str | publishedVersion |
| title | Model generalization experiment. |
| title_full | Model generalization experiment. |
| title_fullStr | Model generalization experiment. |
| title_full_unstemmed | Model generalization experiment. |
| title_short | Model generalization experiment. |
| title_sort | Model generalization experiment. |
| topic | Ecology Space Science Biological Sciences not elsewhere classified Information Systems not elsewhere classified recall rate rise model &# 8217 lower application costs experimental results show complex shape characteristics complex background environments akconv convolution operation global infrastructure maintenance f1 score improvement lightweight network design road crack detection global attention module yolo model based 5 &# 8211 global information c2f module lightweight method automated detection accurate detection yolov8 algorithm yolo achieves regression accuracy quality samples public safety precise solution paper proposes map improvement loss function industrial demands feature redundancy feature information establishing g designed wise cracks dynamically convolutional kernels bounding boxes adaptively adjust accuracy increase 95 increase 9 %, |