| 摘 要: 针对钢丝编织软管表面微小缺陷在强金属纹理背景下隐蔽性强、人工检测效率低的问题，提出一种融合纹理几何先验与轻量化网络的检测框架GAD-MobileNet。首先，利用Gabor滤波器对编织纹理进行方向解耦，并结合DBSCAN聚类算法精准定位并提取“编织纹理基元”；其次，引入轻量级网络MobileNetV4对提取的标准化图像块进行细粒度分类，实现缺陷自动判定。实验结果表明，该方法在自建数据集上的mAP50达到92.8%，检测速度为25.4FPS，在保证高精度的同时满足工业产线实时性要求。该方法有效解决了强纹理背景下微小缺陷特征提取难题，为编织软管自动化质检提供了技术方案。 |
| 关键词: 机器视觉 编织软管 缺陷检测 纹理先验 MobileNet |
|
中图分类号:
文献标识码:
|
| 基金项目: 绍兴市“三位一体”产学研协同攻关专项项目(2024B41002) |
|
| Defect detection method for braided hoses based on GAD-MobileNet |
|
CAICHUNBO, wuzhenyu
|
Zhejiang Sc-i Tech University
|
| Abstract: To address the issues of high concealment of micro-defects on the surface of steel wire braided hoses under strong metal texture backgrounds and the low efficiency of manual inspection, a detection framework named GAD-MobileNet, which integrates texture geometric priors with a lightweight network, is proposed. First, Gabor filters are utilized to perform directional decoupling of the braided texture, combined with the DBSCAN clustering algorithm to precisely locate and extract “braided texture primitives”. Second, the lightweight network MobileNetV4 is introduced to perform fine-grained classification on the extracted standardized image patches, achieving automatic defect determination. Experimental results demonstrate that the mAP50 of the proposed method reaches 92.8% on a self-built dataset with a detection speed of 25.4FPS, meeting the real-time requirements of industrial production lines while ensuring high precision. This method effectively solves the challenge of feature extraction for micro-defects under strong texture backgrounds and provides a technical solution for the automated quality inspection of braided hoses. |
| Keywords: machine vision braided hose defect detection texture prior MobileNet |
