朱超平,杨永斌.基于改进的Faster-RCNN模型的汽车轮毂表面缺陷在线检测算法研究[J].表面技术,2020,49(6):359-365.
ZHU Chao-ping,YANG Yong-bin.Online Detection Algorithm of Automobile Wheel Surface Defects Based on Improved Faster-RCNN Model[J].Surface Technology,2020,49(6):359-365 |
| 基于改进的Faster-RCNN模型的汽车轮毂表面缺陷在线检测算法研究 |
| Online Detection Algorithm of Automobile Wheel Surface Defects Based on Improved Faster-RCNN Model |
| 投稿时间：2020-01-02 修订日期：2020-06-20 |
| DOI：10.16490/j.cnki.issn.1001-3660.2020.06.044 |
| 中文关键词: 汽车轮毂 缺陷检测 深度学习 目标检测 Faster-RCNN |
| 英文关键词:automotive wheels defect detection deep learning target detection Faster-RCNN |
| 基金项目:教育部科技发展中心产学研创新基金项目（2018A02049）；重庆工商大学重点开放项目（KFJJ2019106）；重庆市教育科学规划项目（2018-GX-348）；重庆工商大学自然科学基金项目（1752006） |
|
|
| Author | Institution |
| ZHU Chao-ping | 1.School of Artificial Intelligence, Chongqing Technology and Business University, Chongqing 400067, China; 2.Chongqing Engineering Laboratory for Detection Control and Integrated System, Chongqing 400067, China |
| YANG Yong-bin | 1.School of Artificial Intelligence, Chongqing Technology and Business University, Chongqing 400067, China; 2.Chongqing Engineering Laboratory for Detection Control and Integrated System, Chongqing 400067, China |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 目的 通过构建轮毂在线生产视觉检测系统，预测轮毂生产过程中轮毂表面的缺陷。方法 根据轮毂表面缺陷的定义和评价标准，给出了轮毂表面缺陷的计算模型，采用了改进型的Faster-RCNN目标检测算法，引入了深度生成式对抗网络，消除图像的模糊性，再利用清晰的轮毂表面图像进行模型训练，结合领域专家的判别标准，优化网络参数，构建轮毂表面缺陷检测模型。利用深度学习Pytorch框架，在NVIDIA Tesla P100图像加速卡上进行模型训练，并对模型结果进行对比性实验分析，找出最优的预测模型。结果 在基础网络部分，采用残差模型ResNet101网络比采用VGG16模型的准确率提高了24%。在目标检测网络模型中引入了多通道特征融合模块，准确率提升了2%。再引入FPN金字塔模型，融入低级和高级语义信息，使得输出的多尺度的预测特征图谱效果更好。最后把残差网络的ROI-Pooling算法改为ROI-Align算法，准确率提高了5%。通过对网络模型的不断改进和优化，轮毂表面缺陷的识别率不断提高。结论 利用改进型的Faster-RCNN网络能够识别出轮毂表面缺陷的种类和位置，满足生产环境的要求，具有一定的工程应用价值。 |
| 英文摘要: |
| The work aims to predict the wheel surface defects during the production by constructing a visual inspection system for the online production of the wheel. According to the definition and evaluation criteria of wheel surface defects, a calculation model of wheel surface defects was produced. The improved Faster-RCNN target detection algorithm was used, and the deep generative adversarial network was introduced to eliminate the ambiguity of the image. The limpid image of wheel surface was adopted for model training and the network parameters were optimized combined with the expert's discriminative criteria, thus constructing a wheel surface defect detection model. The deep learning Pytorch framework was used to perform the model training on the NVIDIA’S Tesla P100 image acceleration card, and the model results were analyzed comparatively to find the optimal prediction model. In the basic network, the performance of the ResNet101 network using the residual model improved the accuracy rate by 24% compared with the VGG16 model. The multi-channel fusion module was introduced into the target detection network model, and the accuracy rate was improved by 2%. Then, the FPN pyramid was introduced, and the low-level and high-level semantic information was incorporated, which made the output of multi-scale prediction map better. Finally, the ROI-Pooling algorithm of the residual network was changed to the ROI-Align algorithm, which improved the accuracy by 5%. With continuous improvement and optimization, the recognition rate of the wheel surface defects was improved continuously. The improved Faster-RCNN network can recognize the types and locations of wheel surface defects, thus meeting the requirements of the production environment, and providing great value in the engineering application. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
渝公网安备 50010702501715号