Laser-welding Spots Detection Method Based On Original Orthogonal Moment Values

With the development of society economy, scientific and technological progress,people’s living standard rises ceaselessly, and having a deeper understanding ofhealth，more and more cancer patients hope to reduce the injury to their bodies causedfrom chemotherapy. Therefore, More and more people pay attention to brachytherapysource treatment technology. However, at present, the companies in our country stilluse artificial method to detect the Laser-welding spots of titanium tubes. It will betime-consuming and subjective, besides，it will surely bring injury to workers. Tosolve these problems, this paper introduced a Laser-welding detection method basedon moment values.At present, the researches all over the world about moment theories were almostlimited to invariant moments. However the detect objects in my paper is fixed, so theinvariant moments are inapplicable. According to orthogonal theory and thecomparison results of figures derived from original moment values of Legendremoments, Zernike moments, pseudo-Zernike moments, Fourier-Mellin moments andTchebichef moments of Laser-welding spots, this paper decided to to classify the twoclass Laser-welding spots with original moment values.The paper introduces the concepts and characteristics of moment firstly, andtakes a detailed description of some kinds of important orthogonal moments.Afterwards, the feasibility of the classification with original moment values isvalidated from theory and experiment. Because the input datum dimension is a littlebit too high, this paper conduct kernel principal component analysis to reduce theinput dimensionality and introduce an improved approach. Different from traditionalmethod that keep the principle components with large variances, this paper select theprinciple components with small variances to reduce the inner-class distance of thequalified laser-welding spots.Afterwards this paper select the dimension reduced values of five kinds oforthogonal moments to classify the Laser-welding spots by BP networks, RBFnetworks and support vector machine respectively. The paper make a simpleintroduction to the two kinds of networks firstly, and then use the Hold—Out methodto validate the classification performance by experiments.Finally, the paper make a further validation for the feasibility of the method withsupport vector machine. The paper make a simple introduction to the principle ofsupport vector machine and the related parameter optimization method that geneticalgorithm and particle swarm optimization firstly, and then select the Hold—Outmethod and5-fold cross validation method to validate the classification performance respectively by experiments.The experimental results showed that the Laser-welding spots detection methodbased on original moment values is feasible.