摘要
复杂结构件内部零件装配正确性检测是工业产品检测必不可少的流程之一,但目前仍缺少一种系统稳健性较高的检测方法以完善整个检测流程。针对这个问题,综合计算机断层扫描(CT)检测技术与卷积神经网络分类识别算法,改变以往以连通区域为特征的检测方法,自动识别图像中的感兴趣区域,使合格品的判断标准由区域特征变为个体特征。将CT系统采集的投影数据序列输入卷积神经网络,对工件内部零件进行精确定位并分类,以产品内部零件分类结果作为零件漏装检测的判断标准,以标准工件投影匹配检测工件投影,通过对比完成零件位移检测。通过实验验证可得,所提方法在模拟工件产品和实际产品检测中可完成对工件内部零件漏缺和换位的识别,整个系统对工件内部零件的相互遮挡等因素具有一定的稳健性。
Assembly correctness identification of internal part of complex component is one of the essential processes for industrial product testing. However, there is still lack of a detection method with high systematic robustness to improve the whole testing process. To solve this problem, based on the convolution neural network classification and computed tomography （CT） technology, we propose a detection method to identify automatically the area of interested image, which is different from the detection methods characterized by the connected area in the past. Thus, the judgment criteria of the qualified products is changed from the regional characteristics to individual characteristics. The sequence of projection data collected by CT system is input to the convo｜utional neural network model to precisely locate and classify the internal parts of the workpiece. The result of the internal components classification is taken as the criterion of the detection for the missing parts. The projection of standard workpiece is matched to the projection of the test-workpieee, which can detect the displacement of the parts. The experimental results show that the method can identify missing and misaligned internal parts of the workpiece in the simulation and the experiment. The overall system is robust for the situation such as overlapping among the internal parts of the workpiece.
作者
吴桐
陈平
Wu Tong;Chen Ping(School of Information and Communication Engineering, North University of China Taiyuan, Shanxi 030051, China)
出处
《激光与光电子学进展》
CSCD
北大核心
2018年第4期168-176,共9页
Laser & Optoelectronics Progress
基金
国家自然科学基金(61571404
61471325)
山西省自然科学基金(2015021099)
关键词
图像处理
激光技术
无损检测
装配检测
卷积神经网络
分类识别
角度匹配
image processing
laser technique
non-destructive testing
assembly detection
convolutional neuralnetwork
classification
angle matching
