A non-destructive automatic pavement damage detection scheme based on end-to-end neural networks with multi-level attention mechanism

Yipeng Liu; Chuan Wang; Yingchao Zhang; Xiteng Sun; Cong Du; Dongdong Xie; Yuan Tian

doi:10.1016/j.engappai.2025.111246

A non-destructive automatic pavement damage detection scheme based on end-to-end neural networks with multi-level attention mechanism

Yipeng Liu
, Chuan Wang
, Yingchao Zhang
, Xiteng Sun
, Cong Du^*
, Dongdong Xie
, Yuan Tian^*

^*Corresponding author for this work

Department of Systems Engineering

Research output: Journal Publications and Reviews › RGC 21 - Publication in refereed journal › peer-review

4 Citations (Scopus)

Abstract

The accurate classification and statistics of road damage detection technology are crucial for road condition evaluation and maintenance decisions. However, the accuracy of complex road surface damage detection based on deep learning is still insufficient for real engineering, and even one of the damages may be repeatedly counted. This study develops a new non-destructive automatic road damage detection technology that includes detect road damage based on deep learning and redundant damage image de-duplication. This technology based on multi-level attention mechanism is designed from the perspectives of convolutional kernels and loss functions, improves the accuracy of real road surface damage detection. Compared to the original network, [email protected] and F1 score increase by 5.1 % and 4 % for the public dataset RDD-2020, respectively. This technology achieves de-duplicate accuracy of 94.29 % in the duplicate road damage dataset (DRDD) by adding image processing algorithm, which will accelerate the engineering application of non-destructive automatic pavement damage detection.

© 2025 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

Original language	English
Article number	111246
Journal	Engineering Applications of Artificial Intelligence
Volume	156
Issue number	Part B
Online published	28 May 2025
DOIs	https://doi.org/10.1016/j.engappai.2025.111246
Publication status	Published - 15 Sept 2025

Funding

This work was supported the National Key Research and Development Program of China (Grant No. 2022YFB2602102), the National Natural Science Foundation of China (52408482), Mount Taishan Scholar Young Program of Shandong Province, the City-University Integration Development Strategy Project of Jinan (JNSX2024008). and the Natural Science Foundation of Jiangsu Province (Grant No. BK20230256). The authors gratefully acknowledge their financial support.

Research Keywords

Road damage detection
Deformable convolution network (DCN)
You only look once version 5 (YOLOv5)
Euclidean distance
Attention mechanism

Access Output

10.1016/j.engappai.2025.111246

Other Access Options

Check CityUHK Library

Permanent Link

Right click to copy link

Cite this

@article{9fca06d4d27f498e834c2c75996f0315,

title = "A non-destructive automatic pavement damage detection scheme based on end-to-end neural networks with multi-level attention mechanism",

abstract = "The accurate classification and statistics of road damage detection technology are crucial for road condition evaluation and maintenance decisions. However, the accuracy of complex road surface damage detection based on deep learning is still insufficient for real engineering, and even one of the damages may be repeatedly counted. This study develops a new non-destructive automatic road damage detection technology that includes detect road damage based on deep learning and redundant damage image de-duplication. This technology based on multi-level attention mechanism is designed from the perspectives of convolutional kernels and loss functions, improves the accuracy of real road surface damage detection. Compared to the original network, [email protected] and F1 score increase by 5.1 \% and 4 \% for the public dataset RDD-2020, respectively. This technology achieves de-duplicate accuracy of 94.29 \% in the duplicate road damage dataset (DRDD) by adding image processing algorithm, which will accelerate the engineering application of non-destructive automatic pavement damage detection. {\textcopyright} 2025 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.",

keywords = "Road damage detection, Deformable convolution network (DCN), You only look once version 5 (YOLOv5), Euclidean distance, Attention mechanism",

author = "Yipeng Liu and Chuan Wang and Yingchao Zhang and Xiteng Sun and Cong Du and Dongdong Xie and Yuan Tian",

year = "2025",

month = sep,

day = "15",

doi = "10.1016/j.engappai.2025.111246",

language = "English",

volume = "156",

journal = "Engineering Applications of Artificial Intelligence",

issn = "0952-1976",

publisher = "Elsevier Ltd.",

number = "Part B",

}

A non-destructive automatic pavement damage detection scheme based on end-to-end neural networks with multi-level attention mechanism. / Liu, Yipeng; Wang, Chuan; Zhang, Yingchao et al.
In: Engineering Applications of Artificial Intelligence, Vol. 156, No. Part B, 111246, 15.09.2025.

Research output: Journal Publications and Reviews › RGC 21 - Publication in refereed journal › peer-review

TY - JOUR

T1 - A non-destructive automatic pavement damage detection scheme based on end-to-end neural networks with multi-level attention mechanism

AU - Liu, Yipeng

AU - Wang, Chuan

AU - Zhang, Yingchao

AU - Sun, Xiteng

AU - Du, Cong

AU - Xie, Dongdong

AU - Tian, Yuan

PY - 2025/9/15

Y1 - 2025/9/15

N2 - The accurate classification and statistics of road damage detection technology are crucial for road condition evaluation and maintenance decisions. However, the accuracy of complex road surface damage detection based on deep learning is still insufficient for real engineering, and even one of the damages may be repeatedly counted. This study develops a new non-destructive automatic road damage detection technology that includes detect road damage based on deep learning and redundant damage image de-duplication. This technology based on multi-level attention mechanism is designed from the perspectives of convolutional kernels and loss functions, improves the accuracy of real road surface damage detection. Compared to the original network, [email protected] and F1 score increase by 5.1 % and 4 % for the public dataset RDD-2020, respectively. This technology achieves de-duplicate accuracy of 94.29 % in the duplicate road damage dataset (DRDD) by adding image processing algorithm, which will accelerate the engineering application of non-destructive automatic pavement damage detection. © 2025 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

AB - The accurate classification and statistics of road damage detection technology are crucial for road condition evaluation and maintenance decisions. However, the accuracy of complex road surface damage detection based on deep learning is still insufficient for real engineering, and even one of the damages may be repeatedly counted. This study develops a new non-destructive automatic road damage detection technology that includes detect road damage based on deep learning and redundant damage image de-duplication. This technology based on multi-level attention mechanism is designed from the perspectives of convolutional kernels and loss functions, improves the accuracy of real road surface damage detection. Compared to the original network, [email protected] and F1 score increase by 5.1 % and 4 % for the public dataset RDD-2020, respectively. This technology achieves de-duplicate accuracy of 94.29 % in the duplicate road damage dataset (DRDD) by adding image processing algorithm, which will accelerate the engineering application of non-destructive automatic pavement damage detection. © 2025 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

KW - Road damage detection

KW - Deformable convolution network (DCN)

KW - You only look once version 5 (YOLOv5)

KW - Euclidean distance

KW - Attention mechanism

UR - https://www.webofscience.com/wos/woscc/full-record/WOS:001500457700007

U2 - 10.1016/j.engappai.2025.111246

DO - 10.1016/j.engappai.2025.111246

M3 - RGC 21 - Publication in refereed journal

SN - 0952-1976

VL - 156

JO - Engineering Applications of Artificial Intelligence

JF - Engineering Applications of Artificial Intelligence

IS - Part B

M1 - 111246

ER -

A non-destructive automatic pavement damage detection scheme based on end-to-end neural networks with multi-level attention mechanism

Abstract

Funding

Research Keywords

Access Output

Other Access Options

Permanent Link

Other Files and Links

Fingerprint

Cite this