Recognition Method Of Metal Fracture Images Based On Grouplet And Relevance Vector Machines

This dissertation was supported by the National Natural Science Foundation of China(No.51261024, 51075372)and Guangdong province key laboratory of digital signal and image processing technology(No.2014GDDSIPL-01), Based on the deficiencies of the traditional metal fracture image recognition method, a new metal fracture image recognition method based on Grouplet and RVM is proposed. Compared with the previous research method, the proposed method is superior in recognition rate and speed. The main content of this paper is the following aspects:1. Introduced the principle of Haar wavelet transform, and discussed the theory and algorithms of Grouplet transform on this basis. In addition, analysis Grouplet transform and wavelet transform in terms of the ability to handle images with the example of metal fracture image. The results of the experiments proving that the Grouplet transform show a clear advantage compared to other transforms. Grouplet transform is a new orthogonal transformation based on image geometric flow optimal sparse representation, which can maximize the use of the geometric features of the image. This chapter is the theoretical basis of the full text.2. Combing Grouplet transform with RVM, a recognition method based on Grouplet-RVM is proposed. In the proposed method, Grouplet average energy, Grouplet harmonic entropy and Grouplet kurtosis is respectively used as the image’s feature, RVM as a classifier. The experiment result shows that the proposed method is very effective. Because the Grouplet average energy is proportional to the square of Grouplet coefficient, Grouplet kurtosis is proportional to the fourth power of Grouplet coefficient. Therefore Grouplet kurtosis is more sensitive to the texture change of metal fracture image than Grouplet average energy. Compared with the Grouplet average energy, the Grouplet kurtosis can be more effective to reflect small change of texture feature, so the Grouplet kurtosis is more suitable to extract the feature of metal fracture image. Compared with the wavelet-RVM recognition method, The proposed method can overcome the information of finite directions only obtained by the wavelet-RVM recognition method, and can have a satisfactory recognition rate. Compared with the Grouplet-SVM recognition method, The number of the support vectors based on Grouplet-RVM much less than Grouplet-SVM, what’s more, Grouplet-RVM can use kernel function arbitrarily. The proposed method and Grouplet-SVM recognition method have the same good recognition rate. However, in the training seed, the Grouplet-RVM recognition method is obviously superior to the Grouplet-SVM recognition method, especially in the increase of training samples.3. Variational relevance vector machines(VRVM), is a kind of improved relevance vector machine, which changes the complicated convolution integral operation in relevance vector machine into the easy logarithm operation, thus reduces the amount of calculation of RVM greatly. Combing Grouplet transform with RVM, a recognition method based on Grouplet-VRVM is proposed. In the proposed method, Grouplet average energy, Grouplet harmonic entropy and Grouplet kurtosis is respectively used as the image’s feature, VRVM as a classifier. Compared with Grouplet-RVM, Grouplet-VRVM recognition method is faster in training speed, especially in the increase of training samples.4. In recognition method based on Grouplet-RVM, selection of kernel function parameter is crucial, if choose undeserved, it will make greatly affect on the recognition effect, and for the selection of kernel function parameter, at present, also have no choice basis, aiming at this problem, three kinds of Wavelet kernel function is proposed(Mexican hat Wavelet kernel function, Morlet Wavelet kernel function and DOG Wavelet kernel function), the advanced RVM named Wavelet relevance vector machine(Wavelet RVM, WRVM). Combined WRVM and Grouplet transform, proposed a Grouplet-WRVM recognition method, and applied to the fracture image recognition. The experimental results showed that the Grouplet-WRVM and Grouplet-RVM identification methods are satisfied the recognition rate. However, Grouplet-WRVM recognition method avoids the selection of kernel function parameters, with greater flexibility and adaptability. Compared with Grouplet-RVM recognition method, Grouplet-WRVM recognition method still has a certain advantage.