The Application Of Digital Image Processing To Metallographic Analysis

The proposal of"made in China 2025"greatly promotes the development of Chinese manufacturing industry.Independent research and development of high-precision mechanical equipment has become a major strategic goal of China,and the characteristics of various metal materials have a great impact on the safety and reliability of mechanical equipments.Therefore,as the most widely used and effective technology in the field of metal materials science,the technology of metallographic analysis is been paid more and more attention.Traditional manual metallographic analysis has the defects of heavy work and low efficiency.However,computer aided quantitative metallographic analysis can cover the shortage of artificial metallography and achieve automatic analysis.Metallographic image processing is a key part of the metallographic analysis system.Further research on the integration technology of digital image processing and metallographic analysis has a great economic value for the development of manufacturing industry.The main contributions of this thesis are as follows:Firstly,in this paper,an image denoising method is proposed based on principal component analysis with Gaussian mixture models clustering on the basis of two-stage image denoising by principal component analysis with local pixel grouping.At first,Gaussian mixture models are trained with a series of natural clean images to obtain the parameters,which guide the clustering of noisy patches of the input noisy images,Mahalanobis distance is chosen to measure the similarity of patches instead of Euclidean distance.Then,clustering results are taken as the training data of principal component analysis,the noise will be removed on the PCA domain.Finally,the denoising image is reconstructed.Experimental results illustrate that,in comparison with LPG-PCA,PND-NLM,anisotropic diffusion method,GMM-PCA has a better denoising performance and visual effect,and it preserves texture and edge on the images with rich texture characteristics more completely.Then,the two-dimensional exponential gray entropy thresholding method based on gray scale-gradient histogram is proposed.Firstly,the method of region division based on gray scale-gradient histogram is given and the formulae for threshold selection based on two-dimensional exponential gray entropy thresholding method are derived.Then,the recursive algorithm is adopted to simplify the redundant computation of the intermediate functions.An artificial bee colony algorithm improved by tent chaos mapping is used to search the optimal threshold and reduce solution time significantly.The experiments results on a large number of metallographic images indicate that the proposed method can preserve the objective border accurate,details and features clear.The speed and the performance of segmentation are better than two-dimensional exponential entropy thresholding method based on gray scale-gradient histogram and particle swarm algorithm,two-dimensional oblique exponential entropy thresholding method based on chaotic particle swarm algorithm,also has much improvement on running speed.And then,a grain boundary reconstruction method of metallographical image based on active contour model and mathematical morphology is proposed.Firstly,₁L norm is introduced into the CV model,which combined with local binary fitting,and the normalized proportionality coefficient of global variance and local variance is used to regulate the evolution of contour curve,and curve is driven to close to the real grain boundary,and the effects of precipitated phase in grain boundary for grain boundary reconstruction are reduced.Then,the close operation is applied to the image,and the Rosenfeld method is used to thin the grain boundary in the image after close operation,and the grain boundary with multi-pixels width turns into the closed grain boundary with single pixel width.Finally,the redundant branches of grain boundary are cut off by looking up the given templates.The experimental results indicate that the grain boundary reconstructed by the proposed method is more complete and clear,compared with available grain boundary reconstruction method,and has a better visual effect,which is more competitive.Next,a feature extraction method for metallographic images based on improved Weber local descriptor is proposed.Firstly,anisotropic diffusion based on PM-AOS model is used to filter image and decrease the noise generated by differential excitation.Then,rotation-invariant uniform local binary pattern is used to replace original orientation,and the spatial structure information of the texture in metallographic image is utilized fully.Finally,the histogram of the Weber local feature is constructed to extract the feature vectors of images which are put into a classifier to train.Contrast experiments among the classical texture feature operator local binary pattern,gray level co-occurrence matrix,and original Weber local descriptor,averall accuracy and Kappa coefficient are utilized to measure the classification performance of different algorithms.The results of classification experiments under the DoITPoMS metallographic image database and Brodatz texture database show that,the improved Weber local descriptor proposed by this paper has a higher classification accuracy,which can describe the texture information of the metallographic image accurately.Finally,a metallographic image classification method of cast iron based on unbiased finite impulse response?UFIR?moments and alternating decision forests?ADF?is proposed.Firstly,the UFIR moment eigenvectors of different cast iron metallographic images are calculated,and the training sample matrix is formed and trained in ADF.ADF based on the breadth-first growth mode calculates the weighted entropy of the expected distribution of different sample classes to measure the contribution of the samples.Finally,the classified image is identified according to the trained ADF classifier,and the results of the metallographic image classification of cast iron are finally obtained.The experimental results show that,compared with the proposed image classification methods in recent years,such as the method based on Hu moment with SVM,the method based on Zernike moment and SVM,the method based on Krawchouk moment with SVM and the method based on artificial neural network,the method proposed in this paper has higher classification accuracy,and can identify iron categories quickly and accurately,and improves work efficiency.