摘要
针对传统机器视觉与图像处理中带孔复杂工件缺陷检测存在分类识别模糊和缺陷定位效率低下的问题,提出了一种改进多尺度Faster-RCNN的带孔工件缺陷检测技术。首先,进行带孔工件图像采集与预处理,改进数据集中数据增强方式保证数据训练集质量;其次,搭建Faster-RCNN网络作为基本框架,以ResNet50为主体特征提取网络改进内部残差结构,加强RPN网络提取多尺度特征图数据;然后,改进原有非极大值抑制算法,使用Soft-NMS算法对密集孔洞与缺陷特征分别进行标定分类;最后,进行对比试验将Faster-RCNN与改进多尺度算法比较。结果表明:改进多尺度Faster-RCNN对于工件孔洞特征与缺陷识别精度达到了92%,平均精确度较原算法提升了4.35%,能够同时鉴别工件孔洞特征与其附近缺陷,神经网络鲁棒适应性高。
Aiming at the problems of fuzzy classification and low efficiency of defect location in traditional machine vision and image processing,an improved multi-scale Faster-RCNN method for defect detection of workpiece with holes was proposed.Firstly,the image acquisition and preprocessing of perforated workpiece are carried out,and the data enhancement method is improved to ensure the quality of the data training set.Secondly,the Faster-RCNN network is built as the basic framework,and ResNet50 as the main feature extraction network is used to improve the internal residual structure and strengthen the RPN network to extract multi-scale feature map data.Then,by improving the original non-maximum suppression algorithm,Soft-NMS algorithm is used to classify the dense holes and defect features respectively.Finally,a comparative experiment is carried out to compare the Faster-RCNN algorithm with the improved multi-scale algorithm. The results show that the improved multi-scale Faster-RCNN can achieve 92% identification accuracy for workpiece hole features and defects,and the average accuracy is improved by 4. 35% compared with the original algorithm. It can simultaneously identify workpiece hole features and nearby defects, and the neural network has high robust adaptability.
作者
高沛
王宗彦
段宏伟
GAO Pei;WANG Zong-yan;DUAN Hong-wei(Shanxi Polytechnic College,Taiyuan,Shanxi,030006,China;School of Mechanical Engineering,North China of University,Taiyuan,Shanxi,030051,China)
出处
《建材技术与应用》
2024年第6期10-15,共6页
Research and Application of Building Materials
基金
山西省重点国际科技合作项目(201903D421015)。
