| 摘 要: 针对车载导航导光板表面缺陷像素值分布不均且普遍较小、背景复杂多变等特点，提出了基于改进掩膜区域卷积神经网络(Mask Region-based Convolutional Neural Network, Mask R-CNN)模型检测车载导航导光板表面缺陷的检测方法。首先，引入PinFPN模块改进原有Mask R-CNN的特征融合网络，充分利用高低语义信息构成各级语义、位置信息兼备的共享特征层，提升整体网络的检测精度；其次，通过引入跳层连接结构和SE(Sequence and Excitation)模块对网络的分割分支进行改进，改善了传统Mask R-CNN网络语义信息获取不充分的问题；最后，通过在自建的车载导航导光板数据集上的一系列实验对比，证明了本方法在检测精度和分割上的优势，在自建数据集上的检测准确率达到了95.3%，满足工业检测的要求。 |
| 关键词: 缺陷检测；深度学习；Mask R-CNN；多尺度融合；SE模块 |
|
中图分类号: TP273
文献标识码: A
|
| 基金项目: 浙江省公益性技术应用研究计划项目(LGG18F030001,GG19F030034). |
|
| Research on Surface Defect Detection of Vehicle Navigation Light Guide Plate based on Deep Learning |
|
WANG Hao, LI Junfeng
|
(School of Mechanical and Automatic Control, Zhejiang Sci-Tech University, Hangzhou 310018, China)
2426028009@qq.com; ljf2003@zstu.edu.cn
|
| Abstract: Aiming at the uneven and generally small pixel value distribution and changeable background of surface defects of vehicle navigation light guide plate, this paper proposes a detection method for surface defects of vehicle navigation light guide plate based on improved Mask Region-based Convolutional Neural Network (Mask R-CNN) model. Firstly, PinFPN module is introduced to improve the feature fusion network of the original Mask R-CNN, and high and low semantic information is fully used to form a shared feature layer with both semantic and location information at all levels, so to improve the detection accuracy of the overall network. Secondly, the introduction of skip connection structure and SE (Sequence and Excitation) module improves segmentation branches of the network and insufficient acquisition of semantic information in traditional Mask R-CNN network. Finally, a series of comparative experiments are performed on the self-built data set of the vehicle navigation light guide plate, which proves that the proposed method has the advantages in detection accuracy and segmentation. The detection accuracy on the self-built data set reaches 95.3%, which meets the requirements of industrial detection. |
| Keywords: defect detection; deep learning; Mask R-CNN; multi-scale fusion; SE module |
