Research On Automatic Segmentation And Rating Of Metallographic Image Based On Fully Convolutional Neural Network

Metal materials are widely used in the construction industry and their quality has a great influence on the engineering construction.Major accidents caused by material quality problems not only endanger national safety,but also seriously affect economic development.As the most basic and widely used metal building materials,the quality of steel largely determines the quality of engineering.Therefore,in order to avoid accidents,it is particularly important to test the performance of steel before delivery.Grain size is an important criterion,the performance of steel under the microscopic,the surface grain size of the steel is similar to the cell tissue section of granular distribution image of the area of the grain size,length,and the characteristics of grain number per unit area,etc.all affect the strength,plasticity and toughness properties of steel,and the grain size characteristics in metallographic testing through different metallographic level to reflect.At present,the most commonly used method for the evaluation of metallographic grade is the manual evaluation based on experience.On the basis of digital image processing and deep learning technology,this paper studies the methods of grain boundary extraction and automatic grading of metallographic image of steel,and develops cross-platform intelligent grading software based on these relevant research theories,which lays a foundation for the future research and development of industrialized intelligent metallographic analyzer.The main research contents and achievements of this paper are as follows:(1)The grain boundary data set of the metallographic image of steel and the corresponding metallographic grade data set conforming to the national standard format were established.In this study,the original metallographic image of steel was collected by metallographic microscope,and the original image was de-noised and grain boundary enhanced by time-domain and frequency-domain preprocessing.Meanwhile,under Tensorflow platform,the standard samples and tags are made into a standard data set in a format convenient for deep learning method invocation using the data compression method.Therefore,the research results from the theoretical and practical aspects to enrich the domestic steel metallographic image data set samples.(2)Based on in-depth analysis of the convolution neural network technology,aiming at complex boundary extraction problem which is key point in metallographic rating,this paper proposes a fully convolution neural network based on multi-scale feature fusion boundary extraction method,the adaptive process of pooling in the progress of Downsampling,the trick of multi-scale feature fusion in the progress of Upsampling and jumper connections protect more grain boundary and grain boundary of incomplete information,at the same time,based on grayscale morphology operations using image logic operations to enhance the grain boundary extraction accuracy,so as to make the steel metallographic image boundary extraction accuracy as high as 93%.(3)The multi-scale feature fusion boundary extraction network and composed of Vgg16 network pattern classification rating cascade network was built under the architecture of cascade network,forming an approximate end-to-end has a deeper converged network to realize the steel at the same time since the metallographic image features extraction and independent rating classification purposes,rating accuracy reached 95%,the overall operation time is around 25 ms,can satisfy the requirement of practical application.At the same time,in order to compare the advantage of the network model validation fusion,based on the effective extraction of grain boundary in using the multi-scale feature fusion subnet,the traditional characteristics of the engineering build feature set to combination of the parameters for the design of manual adaptive KNN and SVM algorithm for steel metallurgical grade assess respectively,the characteristics of the fusion network hidden layer can be found from the extraction and classification rating accuracy of traditional method,there is a big advantage,and greatly reduce the computing time under the GPU acceleration.(4)In order to realize the goal of compact intelligent metallographic analyzer in the later stage,a cross-platform steel metallographic intelligent rating software system was developed.With cross-platform QT development environment and OpenCV vision for the software development kit,as well as open source deep learning development framework Tensorflow as the foundation,has realized the related theoretical framework to the transformation of the software,the whole software system mainly include file module operation,the equipment connection module,image preprocessing module,grain boundary extraction module,a key rating module,at the same time in the engineering and the design idea of modular model left interface for the subsequent software development.