Recognition Method Of Metal Fracture Image Based On Manifold Learning

This thesis is supported by the National Natural Science Foundation of China(51261024) and the open Fund of Key Laboratory of the Ministry of Education ofNondestructive Testing technology (ZD200829003). Give full play to deal withhigh-dimensional, nonlinear data of manifold learning. It gets good results in metalfracture image processing. This thesis is composed by following several aspects:1.The paper illuminates the knowledge of the manifold learning, the basic idea ofmanifold learning, several typical manifold learning algorithm and they areComparative analyzed. The content of this part is the theoretical basis of the entirepaper.2.The paper gray level co-occurrence matrix (GLCM) can extract texture feature ofimage, and locally linear embedding (LLE) method of the manifold learning is aneffective nonlinear dimensionality reduction method. Combined the advantages ofGLCM and LLE and undergoing a rigorous derivation, a recognition method of metalfracture image, so-called GLCM–LLE, is proposed. In comparison with the traditionalGLCM method，experiment results show that the proposed method is superior to thetraditional GLCM method in recognition rate, and it has the practical and effectiveadvantages.3.The paper combining the advantage of wavelet transform with manifold learning,this paper presents a recognition method of metal fracture image which based onWavelet-LLE. In this method, using wavelet transform and manifold learning for featureextraction of metal fracture and processing of nonlinear dimensionality reduction, theresult of low-dimensional data is the input to the nearest classifier’s input data foridentifying. In comparison with the recognition method of metal fracture image whichbased on wavelet transform, experiment results show that the proposed method issuperior to feature extraction method of metal fracture image based on wavelet, Theproposed method effectively dealt with extracted features by wavelet, and exploitlow-dimensional geometric distribution structure of the data fully. Thereby, the rate andspeed of recognition are improved greatly by effective dimensionality reduction4.The paper The base of grouplet transform changes according to the image texturefeature, and locally linear embedding (LLE) method of the manifold learning is aneffective nonlinear dimensionality reduction method. Combining the advantage of Grouplet transform with manifold learning, this paper presents a recognition method ofmetal fracture image which based on Grouplet-LLE. In this method, using Grouplettransform and manifold learning for feature extraction of metal fracture and processingof nonlinear dimensionality reduction, the result of low-dimensional data is the input tothe nearest classifier’s input data for identifying. Experiment results show that theproposed method is effective.