Research On A New Method For Automatic Metallographic Rating Based On Ensemble Learning

With the development of modern society,the demand for metal materials is increasing,and the quality requirements are getting higher.Thus it is important to detect the performance of metal materials.The grain size of metal as an significant index has a decisive influence on the performance of metal materials.The prepared samples of metal material is observed and analyzed through high power metallographic microscope with traditional method,which has disadvantages of subjectivity and low efficiency.This paper proposed an algorithm which can automatically measure the grain size of metallographic images from the perspective of image classification under the ensemble learning framework.The main research and achievements are as follows:(1)Metallographic image database has been constructed based on 20 CrMnTi,CF and 55 steel,and preprocessing algorithms were studied.By image screening,manual labeling,unified naming and label setting,nine levels of metallographic image databases have been established.For each image,the grayscale conversion,filtering,sharpening and other preprocessing algorithms have been analyzed according to the evaluation indexes such as MSE,PSNR,and SSIM.It was verified that Gaussian low-pass filtering and ideal filtering sharpening are better than other algorithms in processing metallographic images,which can eliminate noise and sharpen the edge of grains,it provides a solid foundation for the subsequent feature extraction(2)In order to solve the problem of simplification with traditional feature extraction methods,.Based on the ensemble learning framework,a multi-feature set of the texture structure of metallographic images was established,and the influence of four feature selection methods on the reduction of the multi-feature dimension was studied.According to the characteristics of metallographic images which have a large number of random and complex textures,a variety of global and local texture feature extraction methods were studied to achieve the description of multiple texture feature sets.At the same time,considering that the dimension of the feature vector is too large and the number of features is unbalanced between different features,four feature selection methods were studied to form an effective feature space expression,including chi-square test,information gain,random forest method and sequence forward search.(3)Considering that a single classifier cannot match the description of multiple feature sets,the idea of homogeneous and heterogeneous ensemble learning has been proposed.For metallographic images,the integration of feature-fixed heterogeneous classifiers and variable-size homogenous classifiers were proposed based on ensemble learning,by studying the influence of multiple texture features and classifier diversity,the efficiency and accuracy of automatic metallographic image rating was improved.(4)In order to solve the problems of large workload and low accuracy with traditional machine learning classification algorithms.A new multi-level ensemble learning algorithm was proposed based on the convolutional neural network.the convolutional neural networks was applied to automatic metallographic image rating tasks.and for small metallographic images set,the transfer learning was applied to extract multi-level features.On that basis,Random forest algorithm was also introduced based on multi-level features,which has achieved the best effect of metallographic image rating.Finally,the experimental results showed that for most metal materials,the homogenous and heterogeneous integrated classifiers proposed in this paper can achieve a fast rate of 120 ms/frame with rating accuracy of 94.5% for small data sets and low-dimensional features.In addition,for large data sets,Multi-level ensemble learning algorithm based on convolutional neural network was researched,which can achieve 465ms/frame rating speed with accuracy of 98.89%.The two algorithms both have advantages.study of This paper promotes the improvement of rating algorithms,and provides research directions for the subsequent detection of metal materials.