Research On Registration Of PCB CT Images Based On Feature Point

Printed Circuit Board (PCB) is the core component of electronic equipments, andNon-Destructive Testing (NDT) of PCB has an important social, economic and military value.Cone Beam Computed Tomography (CBCT) imaging technology which can obtainhigh-resolution PCB images provides a new way for NDT of PCB.However, there are two inevitable problems in the CT imaging process. First, there is abalance between a wide range of imaging field of vision and high-resolution. To obtainhigh-resolution images of a large-size PCB, sub-regions of PCB need to be imaged separately,and then the obtained images of sub-regions are mosiacing to get the whole high-resolution PCBimages. Second, due to the impacts of leads and other metal substances in PCB, PCB imagescontain many obvious directional artifacts. As a result, the fusion of these images which need tobe obtained under different imaging directions has a significant effect on enhancing the imagequality. The key technology to solve these two problems is the accurate registration of thehigh-resolution PCB images. Therefore, researches on automatic PCB image registrationtechniques are of great significance in speeding up the efficiency of NDT of PCB, improving theimage quality and the level of automation.Towards the characteristics of PCB image, the main content of this thesis is imageregistration techniques based on feature point, which is containing key technologies of detection,descriptor and match of feature point. The main works of this thesis are as follows:Aiming at the problem of poor detection repeatability which is owing to the impacts ofmetal artifacts in PCB images, an improved SIFT detection algorithm is proposed toautomatically extract feature points of PCB images. According to the characteristics of thedistribution of artifacts in PCB images, the SIFT algorithm is improved on three aspects:simplifying the scale space, detecting the minimum points and restraining edge points with scaleadaptive Harris criteria. The experimental results show that the improved SIFT algorithm notonly improves the repeatability and the matching accuracy significantly, but also reduces thecomplexity of algorithm.Since there are severe metal artifacts and numerous similar regions in PCB images, thefeatures can’t be effectively described and causes the match fail. To solve the above problem, afeature descriptor based on the stable regions is presented. First, the proposed method introducesthe Maximally Stable Extremal Region (MSER) detection algorithm to extract the stable regions,then divides the neighborhood based on the log-polar coordinates in a larger neighborhood,which is measured of feature point scale, finally counts the distribution of stable regions andforms a49-dimensional feature descriptor. The experimental results show that the descriptor hasthe ability to adapt to deformation and distinguish similar features, and is able to counter theinfluence of metal artifacts effectively.To solve the mismatch problem that caused by many similar features in the feature matchprocess, a point-pair matching algorithm based on the constraint of directions is proposed. This method firstly generates the point-pairs to suppress the local ambiguity, then removesmismatches by point-pair classification match, constraints with direction and location, finallygives the direction constraints condition between matches and proposes a suppression algorithmof mismatching based on the condition. The experimental results show that the proposedapproach both improves the number of correct matches and the repeatability sufficiently.Aiming at the time-consuming problem for large-size image registration, parallel computingtechnique based on Compute Unified Device Architecture (CUDA) implemented on GraphicsProcessing Unit (GPU) is used to accelerate the process of image registration. This thesisanalyzes the parallelism of detection and match process of the registration algorithm, and findsout the two parts are speedup based on the CUDA programming model. The experimental resultsshow that compared to the CPU, GPU acceleration of S-SIFT and the match algorithm bothoperates efficiency up to10times.