View in EDS PDF Full Text

For Accurate yet Lightweight Road Defect Detection: An Efficient Architecture and a Task-Aligned Training Strategy.

Saved in:
Bibliographic Details
Title: For Accurate yet Lightweight Road Defect Detection: An Efficient Architecture and a Task-Aligned Training Strategy.
Authors: Zhao, Jinchi1 232085400142@stu.ustl.edu.cn, Yu, Yang1 222085400542@stu.ustl.edu.cn, Hui, Yinglun1 243208540506@stu.ustl.edu.cn, Qu, Hening1 15042470716@163.com, Li, Yiran2 lkdzcn@ustl.edu.cn, Zhang, Chunna2 320073500077@ustl.edu.cn
Source: Engineering Letters. May2026, Vol. 34 Issue 5, p2054-2063. 10p.
Subjects: Detection algorithms, Data transformations (Statistics), Intelligent transportation systems, Object recognition (Computer vision)
Abstract: In view of the shortcomings of the existing road defect detection algorithms, such as large number of parameters, high computational complexity, insufficient detection accuracy, as well as false detections and missed detections of minor cracks, this study proposes a lightweight improved YOLOv10n model - YOLOv10n-STP. This improved algorithm features three key improvements: The DySample dynamic upsampler improves the nearest neighbor interpolation upsampler in YOLOv10, enhancing the feature fusion capability of the neck network and reducing model computational complexity without compromising detection accuracy. A lightweight LSCD detection head improves efficiency and accuracy, while shared convolutions reduce parameters and enhance feature fusion, ensuring accuracy in highly complex road defect detection. To make the matching quality and loss weights of different ground-truth prediction pairs more discriminative, we implement a custom PowerTransformer using power transformations to significantly enhance the weights of high-quality prediction boxes, suppressing the influence of low-quality prediction boxes and allowing the model to focus more on high-quality prediction boxes during learning. Combined with the Segmented Power Transformation Adaptive Power Transformation (SPT-APT) strategy, this algorithm improves detection accuracy in road defect detection. Experimental results on the RDD2020 dataset demonstrate that the proposed YOLOv10n-STP model improves the mAP50 accuracy from 52.5% (baseline YOLOv10n) to 55.4%, a gain of 2.9 percentage points. This enhancement significantly strengthens the detection capability for typical road defects such as cracks and potholes, providing an effective solution for intelligent transportation infrastructure inspection. [ABSTRACT FROM AUTHOR]
Copyright of Engineering Letters is the property of International Association of Engineers (IAENG) and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:In view of the shortcomings of the existing road defect detection algorithms, such as large number of parameters, high computational complexity, insufficient detection accuracy, as well as false detections and missed detections of minor cracks, this study proposes a lightweight improved YOLOv10n model - YOLOv10n-STP. This improved algorithm features three key improvements: The DySample dynamic upsampler improves the nearest neighbor interpolation upsampler in YOLOv10, enhancing the feature fusion capability of the neck network and reducing model computational complexity without compromising detection accuracy. A lightweight LSCD detection head improves efficiency and accuracy, while shared convolutions reduce parameters and enhance feature fusion, ensuring accuracy in highly complex road defect detection. To make the matching quality and loss weights of different ground-truth prediction pairs more discriminative, we implement a custom PowerTransformer using power transformations to significantly enhance the weights of high-quality prediction boxes, suppressing the influence of low-quality prediction boxes and allowing the model to focus more on high-quality prediction boxes during learning. Combined with the Segmented Power Transformation Adaptive Power Transformation (SPT-APT) strategy, this algorithm improves detection accuracy in road defect detection. Experimental results on the RDD2020 dataset demonstrate that the proposed YOLOv10n-STP model improves the mAP50 accuracy from 52.5% (baseline YOLOv10n) to 55.4%, a gain of 2.9 percentage points. This enhancement significantly strengthens the detection capability for typical road defects such as cracks and potholes, providing an effective solution for intelligent transportation infrastructure inspection. [ABSTRACT FROM AUTHOR]
ISSN:1816093X