Investigation And Application Of Online High-speed Visual Inspection And Precise Control

Automatic inspection based on machine vision technologies is a research area recentlydeveloped, where various descriptive parameters are extracted from an object image.These parameters can be used to understand the image whilst making certaindecisions. Machine vision techniques have advantages such as non-contact, highaccuracy, wide application and high degree of automation. Therefore, these techniquescan be applied to actual inspections, measurements and control in industry. Sincemachine vision based detection systems are part of the whole production assemblyline, the detection speed must keep highly synchronized with the production assemblylines. The key to implement the synchronization is to develop a high speed detectionalgorithm for camera based control. To promote national economic development, it isvery important to developing an intelligent on-line detection device with our ownintellectual property rights to meet the increasing requirement of the current domesticmarket of intelligent manufacturing.In this thesis, we discuss the key techniques of on-line inspection with regards to thesystem compositions and software packages. As an example of the practicalapplications, on-line crown cap detection is particularly introduced in the thesis. Themajor steps of the detection system are as follows:First of all, we design a machine vision based on-line detection system based on thetheory of the two layers of network control, which provides a new perspective forhigh-speed object detection. In this case, the on-line detection architecture is dividedtwo layers, high layer and local layer. The high layer is used to process the image, andthe local layer is used to control the system. This structure is used to solve theproblem of computational complexity. The system is designed for easy installation,debugging, maintenance and extension.Secondly, a novel control method for the acquired images is proposed, whichcombines the iterative learning control with Kalman filtering. The model’sconvergence and fluctuation range are also derived and analyzed theoretically. Thenumerical simulation and actual experimental results are presented afterwards.Thirdly, the discrete path level set method is proposed, based on the local energy ofimage edges. In this case, the search paths edges are reduced to a few lines. So, thesearch space is greatly reduced. Since both inside and outside of the level set local energy factors are considered, the interference caused by different illuminationconditions is significantly reduced.Fourthly, the circular region projection histogram is used as rotation invariant features,whilst changes2D data into1D vectors and thus the matching efficiency is improved.Meanwhile, a sparse representation method is developed for the rotation matching andflaw detection. In this case, the standard data dictionary is established throughlearning standard samples before the real-time detection occurs. The computation inthe process of detecting is reduced. Therefore, these methods are the keys to achieve areal-time online detection.Fifthly, an experiment platform for the simulation of the production line is designedand built. Then, the results of the simulation and testing at various cases can beobtained by adjusting the related parameters of the experimental platform in the lab.Finally, a crown cap online detection system is developed for practical production.Through10months trial run, the system has reached the requirement of on-line crowncap inspection and the feasibility of the theoretical research and effectiveness havebeen verified.The research results reported in this thesis are not only confined to the crown caponline detection but can also be extended to other products in the field of on-linedetection. It shows a broad market prospect.