Research On Image Segmentation Of Metallographic In GCr15 Bearing Steel Based On Deep Learning

GCr15 bearing steel is one of the most widely used high carbon chromium bearing steel which has been applicated in industry,agriculture,national defence and other fields.The microstructure of GCr15 bearing steel plays a decisive role in its mechanical properties.Therefore,it is of great engineering significance to research the microstructure of GCr15 bearing steel.In this thesis,the carbide segmentation of microstructure images of GCr15 obtained by quenching heat treatment based on deep learning is carried out to enhance the efficiency of quantitative analysis,expand the access to target structure and improve the accuracy of the results.The main content of this thesis are as follows,(1)The metallographic map of GCr15 bearing steel metallographic structure is taken by scanning electron microscope JSM6700,the specimen are prepared with tapered roller,and the subsequent research data set is found out by adjusted or modifed their dimensions.(2)An improved U-Net network model is proposed based on an introduced attention mechanism,namely,the spatial attention mechanism which is added to the original U-Net model to capture and extract the context dependence of global features.At the same time,asymmetric convolution is introduced to instead of the standard square convolution kernel to improve the model segmentation accuracy without extra hyperparameters.(3)An image segmentation model based on CGAN,being called U-GAN in this thesis,the improved U-Net model based on attention mechanism which mentioned above is used as the generator of U-GAN,VGG16 model as the discriminator of U-GAN,and a weighted loss function and an IOU function are added towards the original loss function of the CGAN model,being composite function of U-GAN,which is employed to perform semantic segmentation of undissolved carbide particles in the metallographic structure.(4)The problems of low resolution,low contrast and blurred edge contours of the metallographic images of GCr15 bearing steel.To test abilities of U-GAN algorithm model,the traditional digital image processing methods and semantic segmentation networks are applied to the metallographic images for comparative experimental.The results suggest that the proposed U-GAN acquired the highest Dice coefficient,F1 mearsure and Precision in the undissolved carbide segmentation task which are 0.9395,0.9522,0.9467,respectively.(5)Based on the proposed segmentation image of the U-GAN network model,the content of undissolved carbide particles and the proportion of undissolved carbide particles area in GCr15 bearing steel are quantified.The results show that based on U-GAN algorithm proposed in this thesis,the undissolved carbide particle segmentation of the metallographic structure of GCr15 bearing steel has greatly improved the accuracy comparing with traditional digital image processing methods,it is more accurate and quickly can obtain the average area of carbide in the quenching heat-treated state.