Research On Part Placement Inspection System Based On Machine Vision

With the rapid development of China’s economy,China is facing the challenge of transforming from a big manufacturing country to a strong manufacturing country,in which the intelligent transformation of the manufacturing industry is particularly important.Machine vision technology,as an indispensable part of intelligent production,has been widely used in dimension measurement,defect detection,recognition and grabbing of workpieces.On a factory maintenance line,the workers responsible for disassembly often need to disassemble the product into thousands of small parts and put them into boxes by hand,waiting for the workers on the assembly line to process and assemble.It is very time-consuming and laborious to detect the correct placement of parts or to count the assembly time required for a part only by human effort.This paper tries to build a machine vision system to solve the above problems,and the research contents are as follows:Based on the analysis of the development history of positioning technology in target detection technology,combined with the accuracy and efficiency requirements of the project,a two-stage classification algorithm structure is proposed in this paper,which combines image segmentation based on digital image processing method with in-depth learning technology.This method combines the excellent performance of digital image processing method in classification task with high segmentation accuracy and deep learning method in a single scene,and is suitable for the part box detection task of this topic.Based on the actual production situation of the workstation,cost and functional requirements,the hardware used in the system,such as industrial cameras,lenses,computing equipment,is selected and designed.High resolution black and white cameras are selected according to the small color variability of industrial parts.For the texture and shape of industrial parts are relatively simple and easy to distinguish,a lightweight deep neural network is selected with edge computing equipment.After the selection is completed,the system is set up and debugged to prepare for subsequent data collection and training.In the first phase of the detection task,an adaptive threshold segmentation method was chosen after the experiment,combined with appropriate morphological processing,to split the outline of the part box,and straighten it by perspective transformation.Location and partition of part position in the detection algorithm is completed by recording the position and shape of each part slot in the CAD top view of the part box.Part box images are collected and segmented using the first-stage algorithm.The resulting part images are formed into part datasets,and the brightness,contrast,sharpening and Gauss blurring data are expanded to increase the number of data.Finally,according to the system hardware capabilities and project requirements,two kinds of network structures,Mobile Net V3-Small and Res Net-18,are selected to train in the part dataset.The results show that Mobile Net V3-Small network works better,with 98.87% accuracy and 89 fps detection speed,which meet the system performance expectations.