| 李洪奎,陈浩,刘韵婷,张兴伟,冯欣悦.基于改进自编码器结构的轮胎缺陷检测[J].电子测量与仪器学报,2024,38(10):170-179 |
| 基于改进自编码器结构的轮胎缺陷检测 |
| Tire defect detection based on improved autoencoder structure |
| |
| DOI： |
| 中文关键词: 生成对抗网络 NAM 深度学习 FAM 轮胎缺陷检测 PatchGAN |
| 英文关键词:generative adversarial network NAM deep learning FAM tire defect detection PatchGAN |
| 基金项目:辽宁省自然科学基金（2022-KF-14-02）、辽宁省教育厅面上项目（LJKMZ20220617）资助 |
|
|
| Author | Institution |
| Li Hongkui | School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159，China |
| Chen Hao | School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159，China |
| Liu Yunting | School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159，China |
| Zhang Xingwei | School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159，China |
| Feng Xinyue | School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159，China |
|
| 摘要点击次数: 447 |
| 全文下载次数: 1176 |
| 中文摘要: |
| 针对部分轮胎X光缺陷图像中缺陷对比度较低、缺陷占比较小，导致缺陷难以检测的问题，采用了一种基于生成对抗网络的改进模型，以提高轮胎缺陷的检测精度。首先分析了传统生成器所存在的一些问题，然后以GANomaly作为基础模型，引入了注意力机制模块NAM、流对齐模块FAM和PatchGAN，旨在增强模型的特征提取能力和图像重构能力。注意力机制模块NAM通过归一化处理增强了模型对缺陷区域的关注度，流对齐模块能够将低分辨率特征图中的特征点精确地映射到高分辨率特征图的对应位置，从而确保多尺度特征之间的信息一致性和有效融合，而PatchGAN则通过局部判别器增强了模型对局部特征的识别能力。为了验证改进模型的有效性，在相同的自制数据集上对4种轮胎缺陷类型X光图片进行测试。测试结果表明，改进后的模型在受试者工作特征曲线面积（AUC）和平均精度（AP）两个关键指标上均取得了显著提升，AUC值达到了96.4%，AP值达到了95.8%。这些结果表明，改进后的模型有效增强了特征提取和图像重构的能力，提升了缺陷检测的精准度。 |
| 英文摘要: |
| To address the challenges of low contrast and small defect sizes in some X-ray images of tires, which make detection difficult, an improved model based on generative adversarial networks (GANs) is proposed to enhance the accuracy of tire defect detection. Initially, issues with traditional generators are analyzed. Building upon the GANomaly model, the proposed approach incorporates the attention mechanism module (NAM), flow alignment module (FAM), and PatchGAN to enhance feature extraction and image reconstruction capabilities. The NAM enhances the model’s focus on defect areas through normalization, while the FAM accurately maps features from low-resolution to high-resolution feature maps, ensuring information consistency and effective fusion across multiple scales. PatchGAN, with its local discriminator, improves the model’s ability to recognize local features. Validation tests on a self-constructed dataset of four tire defect types demonstrate significant improvements in key metrics, achieving an AUC of 96.4% and an AP of 95.8%. These results indicate enhanced feature extraction and image reconstruction capabilities, leading to improved defect detection accuracy. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
