Research On Microstructure Segmentation Method For Metallographic Image Of Aluminum Alloy

As one of the most widely used alloy materials,aluminum alloy is very important for controlling its performance.Studies have shown that the properties of aluminum alloys are closely related to the type,content and distribution of internal compounds.Experts can analyze and judge the properties of aluminum alloys through metallographic images of aluminum alloys.Therefore,this paper takes the aluminum alloy metallographic image as the research object,and takes the aluminum alloy quality assessment as the background,combined with image processing,machine learning and deep learning methods,to study the microstructure segmentation method of aluminum alloy metallographic image,the main research work as follows:(1)This paper reviews the research status of aluminum alloy metallographic analysis technology,and studies the image feature extraction method,segmentation and classification algorithm.And it also summarizes the difficulties and breakthroughs in the field of metallographic analysis,and lays a theoretical foundation for the design and implementation of subsequent algorithms.(2)A method for evaluating the quality of aluminum alloys based on latent Dirichlet analysis is proposed.The method is applied to model the area,length,shape and distribution of the microstructure of aluminum alloy metallographic images.The latent Dirichlet analysis model is used to simulate the generation process of different quality level images,and the automatic classification of different grade metallographic images is realized.At the same time,the distribution of the main factors affecting the quality of the aluminum alloy sheet can be obtained.(3)An unsupervised segmentation method based on simple non-iterative clustering(SNIC)and improved density-based spatial clustering of applications with noise(DBSCAN)is proposed.The method firstly uses superpixel segmentation to reduce the dimension of highresolution metallographic image,and solves the problem that the image resolution is too large to calculate complex.Then,adaptive DBSCAN clustering is used to calculate appropriate aggregation parameters for each image,and the aggregation is over-segmented.The microstructure of the aluminum alloy eliminates the complicated tuning process.On this basis,a marking method that allows manual correction is provided,which solves the problem of difficult marking of metallographic images and provides good support for supervised learning.(4)A method for semantic segmentation of metallographic images based on improved U-Net network is proposed.Firstly,the data expansion method of random cropping image sub-block and sub-block fusion is designed,which solves the problem that the metallographic image data set is scarce and the image resolution is high.Then shorten the structure of the network and join the Batch Normalization(BN)layer to solve the model fitting problem.Finally,the weighted cross entropy loss function is designed to correct the positive and negative sample imbalance problem in the database and the weight update direction deviation caused by the difficult sample imbalance problem.Experiments show that the improved network can get a good semantic segmentation effect.