面向在线语义分割机器视觉检测鉴别的多任务协同调度方法

A multi-task collaborative scheduling approach for online semantic segmentation machine vision inspection

在线语义分割机器视觉检测鉴别方法综合采用视觉成像技术、语义分割方法,感测被测对象特征与辨别被测对象优劣,在线应用要求鉴别时间短。提出一种面向在线语义分割机器视觉检测鉴别的多任务协同调度方法,首先建立检测鉴别多任务协同调度任务模型,将在线语义分割机器视觉检测鉴别流程抽象为定位、光源控制、相机成像、图像语义分割4种活动,以及高分辨率图像分块分割、分批调度分割2种子活动;其次研究检测鉴别协同调度、高分辨率图像分块调度、多网络多图像分批调度实现方法;最后,在线机箱装配质量检测鉴别系统进行应用试验。初步应用结果表明,单个机箱检测、小批量机箱检测鉴别总时间T_(inspect)分别为5.20 s、24.48 s,分别比典型多任务调度模型缩短34.59%、35.11%,多任务协同调度优化效果明显。方法具有建模简单、实时性好的特点,可拓展用于其它深度学习机器视觉检测鉴别中。

Online semantic segmentation machine vision inspection and identification method using a combination of visual imaging technology,semantic segmentation methods. The system can sense and identify the characteristics of the object under test. The online application of the system requires a short identification time. In this paper,a multitask collaborative scheduling method for online semantic segmentation machine vision inspection and identification was proposed. Firstly,a multi-task collaborative scheduling task model for inspection and identification was established.The online semantic segmentation machine vision inspection and identification process were abstracted into four activities of positioning,light source control,imaging,semantic segmentation,and two sub activities of high-resolution image segmentation and batch image segmentation. Secondly,the implementation of collaborative scheduling for inspection and identification,block segmentation scheduling for high-resolution images,and batch segmentation scheduling for multiple images were studied. Finally,the online chassis assembly quality inspection and identification system is tested for application. Preliminary application results show that,the total inspection time Tinspectfor single chassis inspection and small batch chassis inspection is 5. 20 s and 24. 48 s respectively,which is 34. 59% and 35. 11% shorter than typical multi-task scheduling models,respectively. The multi-task collaborative scheduling optimization is effective. The method has the characteristics of easy modelling and good real-time performance,and can be extended for use in other deep learning machine vision inspection and identification.