Research On Interactive PCB Image Segmentation Based On Graph Theory

Printed Circuit Board (PCB) is an important part of electronic equipments, andNon-Destructive Testing (NDT) of PCB has an important economic, social and military value.Cone Beam Computed Tomography (CBCT) imaging technology which can obtainhigh-resolution PCB images provides a new way for NDT of PCB.To accomplish NDT of PCB, the captured PCB image should be processed and analyzedfirstly which mostly depending on PCB image segmentation technology. PCB which is providedwith complex inside structure consists of thin substrate, conducting wire and vias.Three-dimensional image of PCB which acquired by CBCT usually bears poor quality as a resultof scattering artifacts and metal artifacts. It is very difficult to achieve accurate segmentationresult and also have a bad influence on the accuracy of NDT of PCB.Therefore, researches onPCB image segmentation techniques are of great significance in speeding up the efficiency ofNDT of PCB. Among a large number of image segmentation methods, the graph theory basedmethod develops quickly in recently years. The graph cuts technology which apply graphoptimal partition in image segmentation and formulate image segmentation as an minimizationmethod has achieved many research production.According to the characteristics of PCB image, the main content of this thesis is imagesegmentation techniques based on graph cuts. The main works of this thesis are as follows:Traditional stroke-based graph cuts methods only use the gray information of seeds toupdate the regional term model of object and background which proved to be inefficient andinaccurate. To address the problem the regional term models of object and background aregenarated by gaussian mixture model where gray level of seeds are used to initialize theExpectation-Maximization method and prior information is used to remove errors.Theexperimental results show that the proposed method not only have a better use of hints of user’sinteraction, but also obtain a more stable regional term model.Currently, the state of art graph cuts method is inefficient with the non-uniform PCB image.To solve the above problem, an impoved graph cut model in which a background potential isintroduced is proposed. The scattered data based surface fitting method is used to generate thebackground potential. Through a comparison of different PCB image segmentation experiments,results demonstrate the proposed method has good performance with the non-uniform PCBimage.To avoid the general fluctuation effect of existing seed-based graph cut methods, a new kindof interactive segmentation method named local progressive cuts is proposed. The local energy term is generate adaptively following the seed added by user and is introuced into the graph cutframework as an extra constrain. Through a comparison of different PCB image segmentationexperiments, results demonstrate the proposed method not only avoids the general fluctuationeffect, but also has better performance compared with the state of art method such as graph cutsin terms of segmentation accuracy, controllability, and user experience..The graph cut based image segmentation is time consuming and feedbacks slowly. Theruntime of various part of the method is analyzed. To accelerate the system, a random samplestrategy in training gaussian mixture model and graph reuse straegy are used. Comparing to theorigin algorithm, the runtime of the improved algorithm is reduced to about50%.