Study On Grain Size Rating Method Of Metallography Based On Deep Learning

Grain size is an important criterion for evaluating the performance of steel.From a microscopic perspective,the surface grains of steel exhibit a granular distribution image similar to that of cell tissue slices.The area,length,and number of grains per unit area of the grains all affect the strength,plasticity,and toughness of the steel.These grain characteristics are reflected by different metallographic grades in metallographic testing.At present,the most commonly used method for evaluating the metallographic grade is still to rely on manual experience assessment.The manual determination of the metallographic grade largely depends on the quality of the workers,with a high degree of uncertainty.On the basis of digital image processing and deep learning technology,this project studies the methods of grain boundary segmentation and automatic rating in steel metallographic images.Based on these related research theories,a cross platform intelligent rating software has been developed,laying the foundation for the future development of industrialized intelligent metallographic analyzers.The main work of this article is as follows:1)Acquisition of metallographic images and establishment of datasets.Prepare metallographic specimens in accordance with the national standard "Method for Testing Metal Microstructure".Then the image is grayed and filtered to achieve denoising and histogram equalization to enhance image contrast.Manually annotate the label image in the early stage,and then segment and enhance the data to expand the dataset.Subsequently,a complete dataset suitable for deep learning was produced.2)Research on Grain Boundary Segmentation Algorithms for Metallographic Images.Elaborate on commonly used grain boundary segmentation algorithms.And the actual effects of these algorithms in metallographic tissue images were experimentally and analyzed.Explore the advantages and disadvantages of various algorithms in metallographic images,laying the groundwork for future improvement and development of grain boundary segmentation algorithms suitable for metallographic tissue images.3)Research on Metallographic Image Rating Algorithms.Rating algorithms are divided into two categories: traditional image rating methods and deep learning based image rating methods.In deep learning based image rating methods:(1)The improved FCN network is a convolutional progressive upsampling FCN semantic segmentation method.The activation function and up sampling process in the network are improved.(2)An improved U-Net adopts a dynamic fusion method: in high-order feature maps,a fusion strategy is selected to achieve the goal of adding more feature channels.Choosing a logical addition strategy within a low-level feature map can better preserve feature information.There is a certain improvement effect on the grading of metallographic grain size.4)Implementation of metallographic image rating system.This article studies and develops an intelligent grading software system for metallographic grain size based on deep learning,which achieves automatic detection,segmentation,and grading of grain size in metallographic images.In this system,an improved deep learning semantic segmentation U-Net network was used for pixel level segmentation of grain boundaries in metallographic images,ultimately achieving an accuracy of 97.86% for grain boundary recognition.