The Metallographic Structure Image Analysis Research Of Ferrous Metal Based On Semantic Segmentation

Steel is an essential material in the modernization construction,and the quality of steel can be evaluated through the metallographic analysis.How to make the metallographic analysis of steel efficiently and accurately is of vital importance in the steel production.Meanwhile,in recent years,with the continuous development of Deep Learning and the update of the computers,more and more deep learning algorithms have been introduced into People’s Daily life under the support of stronger computing performance.The main content of this paper is to apply the deep learning segmentation algorithm to the metallographic structure analysis efficiently.The detail is as follows.As to the precise segmentation of small objects in the metallographic structure of iron and steel.This paper proposes a method of multi-path enhancement in feature extraction of the model,It includes convolution residual unit,multiresolution fusion unit,and chain exactly to the residual operation unit,at the same time,we reference SENet and SKNet,we introduced the multi-dimensional channel attention mechanism to fully excavate the information of feature maps in different sizes,and can make the model backbone be better at capturing the detailed feature information of images.We observed the appearance characteristics of several common metallographic structures of ferrous metals and found that the shape information is complex and difficult to distinguish.In order to improve the recognition ability of the model,we adopted two methods to improve the model,Firstly,as inspired by Inception,we introduced the multi-dimensional information to our models.Then,we analyzed the structure of the traditional convolution,and we introduced the deformable convolution to improve the model which could capture the shape information of the metallographic structure of steel better.In the process of segmentation of ferrous metal metallographic structure,The predicted result of the model that there will be another metallographic structure in one kind of structure,which is called "chessboard problem".This phenomenon mainly occurs in the decoding stage of segmentation.Based on the multi-path enhancement method,we improved the model in two ways,One is to make the experiment to verify in the decode stage of the model by introducing sampling with traditional convolution.The other is that we improved the traditional deconvolution by introducing "Pixel Deconvolution" to reduce the influence of "checkboard problem".In real life,the efficiency of the model and memory cost are important factors and should be considered.At present,the model can predict the metallographic structure of steel with high accuracy,however,the inference speed still cannot meet the requirement of real-time.As a result,We combined with model compression to prune and quantify the pretrained model which coulde make the model could make a higih efficiency inference to make real-time segmentation by costing a lower computational resources and memory footprint.