arrow
返回

Resource-Efficient Surface Defect Detection on Edge Devices Using a Hybrid Descriptor Framework with Data-Driven Feature Selection

delete2026-09-02
delete0
delete
OA
AI
B
Burhan Duman
DOI:10.3390/electronics15173915delete
delete原文链接
delete原文求助
delete分享
delete收藏
摘要

摘要

En 中文
针对生物外壳、光伏板和路面基础设施等多样化表面类型的表面缺陷检测对于工业质量控制与结构健康监测至关重要。尽管深度学习(DL)模型在该领域表现良好,但其高计算和内存需求限制了在资源受限边缘设备上的部署。为解决此问题,本文提出了一种计算高效的特征级提取框架,结合局部二值模式(LBP)、Gabor滤波器和离散小波变换(DWT)以捕获表面缺陷的互补纹理、方向和频率特征。AdaBoost驱动的特征选择策略将高维混合特征池缩减为25个最具判别性的属性,并应用SHapley Additive exPlanations(SHAP)分析解释特征贡献。所提框架在EggCrack数据集上实现100%准确率,在更具挑战性的ELPV数据集上达到79.0%准确率。在ELPV上,其优于MCU优化的MCUNet-In0(78.16%),且与EfficientNetV2-B0(83.16%)的差距仅4.16个百分点,同时计算量(FLOPs)减少约7577倍、模型体积缩小89倍。部署在Raspberry Pi 5 CPU上时,框架实现每幅图像37 ms的端到端推理时间,支持实时检测。结果表明,所提混合描述符与选择框架为基于CPU的边缘设备实时表面缺陷检测提供了理想的准确率-效率平衡。
Keyword:
edge intelligence
surface defect detection
hybrid feature fusion
data-driven selection
resource-efficient computing
explainable AI
real-time image processing

期刊

Electronics 封面图
Electronics
IF:
2.6
论文数:
1.1W
被引数:
4.7W

机构

I
Isparta University of Applied Sciences
学者数:
544
论文数: 491
被引数: 471
引用论文

引用论文

Review of Pavement Defect Detection Methods
err2020-01-01
err190
errOAAI
errCao, Wenming; Liu, Qifan; He, Zhiquan
err分享
err收藏
err分享
err收藏
Tire appearance defect detection method via combining HOG and LBP features
err2023-01-12
err0
errOAAI
errHongbin Liu; Xinghao Jia; Chenhui Su; Hongjuan Yang; Chengdong Li
err分享
err收藏
Flexible and Fully Quantized Lightweight TinyissimoYOLO for Ultra-Low-Power Edge Systems
err2024-01-01
err1
errOAAI
errMoosmann, Julian; Muller, Hanna; Zimmerman, Nicky; Rutishauser, Georg; Benini, Luca; Magno, Michele
err分享
err收藏
学者 查看更多内容