Recognition Methods Of Metal Fracture Surface Images Based On Grouplet Transform And Kernel Methods

This paper is funded by the National Natural Science Foundation of China (51261024,51075372) and Guangdong province key laboratory of gddsipl (No. 2014-01) for digital signal and image processing technology. Combining with Grouplet transform and advantages of several typical kernel methods (KPCA, KMSE, KFDA), the paper put forward some metal fracture image recognition methods, Grouplet-KPCA, Grouplet-KMSE, Grouplet-KFDA and Grouplet-GKPCA. And comparative analysis those new methods, and obtained innovative achievements. Paper main research content includes the following sections:The first chapter discusses the research significance of the Grouplet transform combined with kernel method, which is applied to the metal fracture image recognition. The fracture treatment, Grouplet transform and kernel method research status in domestic and abroad are reviewed in this article. Then put forward the research content and innovation of this paper.The second chapter discusses the Grouplet transform theory and algorithm, using the change adaptive ability of Grouplet transform applied to image texture structure, applied the Grouplet transform into the metal fracture in image processing, and combined with kernel principal component analysis (KPCA), metal fracture image recognition which is based on Grouplet-KPCA method is proposed.At the same time, a wavelet-KPCA method is proposed. The experimental results show that the method overcomes the shortcome that the wavelet KPCA identification method can only be limited in the direction of the image information. It achieves a higher recognition rate. Grouplet kurtosis is more sensitive than Grouplet entropy to fracture image texture changes.Particularly, it is suitable for the feature extraction of metal fracture surface, therefore, it recognises more effective than metal fracture feature extraction based on entropy-KPCA Grouplet.The third chapter puts forward image recognition method based on Grouplet-KMSE metal fracture, combining advantages of Grouplet transform and minimum mean square error (KMSE). At the same time, make Grouplet coefficient of the mean, variance and entropy and kurtosis as characteristic vector, and introduce the KPCA algorithm KMSE in the characteristic value of node selection algorithm to get the optimal feature vector for recognition. The experimental results show that the method is effective.the Grouplet transform overcomes the shortcome that wavelet transform can only be limited in the direction of the image information, thus, is better on identification. Contrast with Grouplet KMSE and Grouplet KPCA method, it has higher recognition rate. However, Grouplet-KMSE in feature selection is adopted in this solution, choosing contribute large feature vector is used to identify, therefore Grouplet-KMSE method has a significantly faster recognition speed.The fourth chapter, the KFDA method can make the best of the sample class information, and could retain high-dimensional characteristics information well. So combined the advantages of Grouplet transform and KFDA, Grouplet-KFDA fracture image recognition method is proposed. Grouplet-KFDA fracture image recognition method select node by KPCA and minimum similarity functions. It can reduce the redundant information,and has more rapid and accurate recognition results. Experimental results demonstrate the effectiveness of proposed methods.The new method can get fracture image feature details extraction and retain high-dimensional information.The fifth chapter, in order to overcome the disadvantage that the classic kernel method relies on kernel function too much, A method method has been proposed, called Grouplet-GKPCA. It fabricates kernel function via KPCA method. Grouplet-GKPCA can generate a new kernel function by embedded the training samples into the kernel function, and generate a new kernel function by means of an arbitrary function. So the Grouplet-GKPCA method itself has metal fracture surface image information. In addition, the Grouplet-GKPCA kernel functions can be constructed by arbitrary functions, making the choice of kernel function more flexibly, and need not be limited to the existing kernel functions. Results also show that the method is effective.The sixth chapter summarized the contents of this paper, and proposes directions for further research.