Automatic Grain Size Rating Of Metallographic Based On Deep Learning

In the automatic evaluation of grain size of metallographic structure,the accuracy of grain boundary identification directly affects the evaluation accuracy of grain size of metallographic structure.In traditional image processing algorithms,grain boundary identification depends on the purpose of grain boundary and grain is the obvious boundary between feature to extract the grain boundary,we need the required boundary in the process of the algorithm features requires human to extract,but can extract the characteristics of the artificially is limited and the human mind and indirect limited,Simple features like color differences between grain boundaries and grains,light and shade differences,and gradients between pixels,gradients in the shallow layers.So when dealing with the complex and changeable microstructure image,traditional image processing algorithms is difficult to accurately identify the complex and changeable grain boundary information,therefore,we hope that on deep learning algorithms in different levels of image feature extraction and combining with the advantages,enhance accuracy and velocity of grain boundary identification,and then implement efficient accurate assessment of grain size.In this paper,an automatic grading system of metallographic grain size based on deep learning is proposed,which realizes the automatic detection,segmentation and grading of grain size of metallographic images.In this system,the U-Net network with deep learning semantic segmentation is used to segment the grain boundary of metallographic images at pixel level.Finally,the accuracy of grain boundary recognition is 96.07% on the semantic segmentation network based on U-NET.By adding random boundary missing mask way to expand so as to improve the segmentation accuracy of deep learning network data sets,then according to the eight neighborhood track and the growth of the diffusion method,the good image segmentation binarization operation,for thinning image after binarization processing skeleton operation,through the eight neighborhood method lookup endpoint,remove the pixel length is less than a certain threshold of "burr",Finally,the missing areas of grain boundary are connected to realize the completion of some missing areas.Finally,according to national standards,all closed grains in the metallographic image were counted and rated by the image connected domain extraction algorithm and area method,and the rating report was finally generated.The whole rating process lasted 18 seconds from the image selection by the operator to the rating report generation,and the effect was good.