Research On The Machine Vision Inspection And Perception Key Technologies And Their Applications In Intelligent Manufacturing Equipment

With the forthcoming of Industry 4.0 era,developing smart factory and intelligent manufacturing industry has became the common view around the world.The Intelligent manufacturing process is usually very complex,which consists of many production processes.Every production process is accomplished by one or more intelligent manufacturing equipments.For intelligent manufacturing equipment,environment sensing and intelligent control is the key technology to ensure their highly adaptive,highly precise and highly intelligent operation.It is also the primary problem to be solved in intelligent manufacturing equipment development.However,the traditional sensing and control technology is unable to meet the real-time,highly precise,intelligent,modular,non-contact requirements of intelligent manufacturing equipment.Meanwhile,machine vision measurement and control technology provides an optimal solution to this difficult problem.Hence,machine vision is a hot research topic around the world.To improve the adaptability and intelligence level of intelligent manufacturing equipment,this thesis develops a novel machine vision inspection and control system,which combines robotics and machine vision to improve the flexibility.The main work of this thesis is to solve three difficult problems encountered during the developing of this system,i.e.automatic object loading/unloading,surface defect detection and stack counting.The main works and contributions of this thesis are as follows:(1)A machine vision measurement and control method system for intelligent manufacturing equipment is proposed.The main components of this method system are analysed and surveyed,which can be used as theoretic foundation of intelligent manufacturing equipment design.(2)To solve the object detection and location problems during automatic loading or unloading,a line segment based method is proposed to detect the objects and estimate their pose.The proposed method consists of off-line template learning and on-line object detection,the first step is adopted to construct a pattern from the template image,while the latter recognize the objects and estimate their pose according to the learned pattern.To obtain the template pattern,first a line segment detector is applied to the template image to extract the line segments.Then an edge grouping algorithm is utilized to obtain closed contour based on minimizing a cost function defined on line graph.Finally,length of each line segment in the closed contour and the inter-line angles are used as the template pattern.To detect the objects in realtime image,first the same process is used to extract the closed contours.Then the according length of line segments in the closed contour and the angles are sepa-rately matched with defined similarity functions.Finally,for each detected object,its pose is estimated by contour corner matching algorithm.Our experiments have indicated the proposed algorithm is highly accurate,the average location error is within 3 pixel and the time consumed by one detection process is less than 400ms.It fully meets the requirements of objects detection and location in various industrial applications.(3)To solve the highly accurate quality inspection problem for intelligent manufacturing products,a super-pixel based surface defect detection and segmentation method is proposed,which consists of object location,measuring regions extraction and defect detection in multiple regions.During object location,first an entropy rate clustering algorithm combined with shape fitting algorithm is adopted to segment the inspection object.Then based priorrTegion information,the multiple measuring regions are extracted.For the large measuring regions,a super-pixel segmentation,grouping and high frequency components detection algorithm is proposed to detect the various defects.For the narrow annular measuring regions,a ridge detection algorithm is applied to seek the defects along its projection profiles.The experiments have verified the proposed algorithms are able to detect and segment the various defects across different measuring regions,and the defects detection accuracy is above 99.4%,while the defect segmentation accuracy is around 90%.Besides,the proposed algorithms are robust,it is well suited to various quality inspection applications.(4)To account for the stacked sheets counting problem in product lines,a line-scan imaging system is designed to image the edges of the stack,and two post-processing algorithms are proposed to discern the sheets from the obtained images.One method is based on the similarity between the stack profiles and Gaussian shape,and ridge detection algorithm is utilized to recognize the stacked items.The other method is based on the additive model of stack image,and wavelet decomposition and coefficient shrinkage algorithm is used to extract robust feature for object detection.In the ridge detection based algorithm,to remedy the shortcoming of traditional ridge detector,a bi-Gaussian ridge detector and a ridge-valley descriptor are proposed to enhance the contrast of ridge response between the sheets and the inter-sheet gaps.Within the wavelet based algorithm,an energy distribution based algorithm is proposed to determine the optimal threshold for each level.Based on the obtained features,a sliding window peak detection based item detection and ridge region based verification algorithm are proposed to discriminate the true items in the stack.The experiment results indicate the average error with proposed algorithms is less than 0.03%,and they also demonstrate excellent repeatability.The proposed system is sufficient to meet high accuracy counting requirements for various stacks.(5)To implement the proposed machine vision inspection and control system,a message passing and reference point based control scheme is proposed to make the different parts to work cooperatively.A modular and image processing plan based method is proposed to implement machine vision inspection software,which improve the system development efficiency and application scope.Real operation of the equipment indicates one inspection job can be accomplished within 10s,it can meets the requirements of various manufacturing tasks.In conclusion,this thesis studied a general purpose machine vision inspection and control system,and solved three typical vision perception tasks in the equipment development,including automatic loading/unloading,quality inspection and stacks counting.The proposed algorithms are highly accurate and robust,which have very valuable academic and practical meanings.