Research On Grain Boundary Extraction Algorithm Of TA15 Titaniun Alloy Image Based On U-Net Network

As early as the 1950 s,titanium alloys have become the main materials in the aviation field,and the properties of titanium alloys are closely related to their microstructure.In recent years,with the development of computer technology,the processing of titanium alloy metallographic images has also become a hot field.The premise of studying the properties of TA15 titanium alloy metallographic image material is to quantitatively analyze the image grain boundary α phase of TA15 titanium alloy,and the extraction of grain boundary α phase is the most important step in this process.By analyzing the microstructure characteristics of the grain boundary α phase in the TA15 titanium alloy image,it is found that there are two difficulties in the process of grain boundary extraction:one is that the lamellar structure of the background is very similar to the grain boundaryα phase,which makes it impossible to clearly and intuitively identify the grain boundaryα phase;the other is that the equiaxed α phase is embedded in the grain boundary αphase,making the grain boundary α phase discontinuous.In view of the above difficulties,the main research contents of this paper are as follows.Firstly,in order to solve the problem that the lamella structure is similar to the grain boundary α phase texture,color and other characteristics,which leads to the wrong extraction of the lamella structure,a grain boundary extraction algorithm of TA15 titanium alloy images based on DR-UNet is proposed.Combining the structure of the grain boundary α phase features and human visual perception field changes and multi-scale characteristics,introduce deform convolution and multi-scale feature blocks,reconstruct the encoding network and decoding network of U-Net model;At the same time,short skip connections are introduced in skip connections,and as the number of network layers deepens,the number of convolution blocks used is gradually reduced,the feature fusion is strengthened,and network parameters are stably updated to improve the accuracy of grain boundary extraction.Secondly,a grain boundary extraction algorithm for TA15 titanium alloy images based on DRCM-UNet is proposed for the discontinuity of grain boundaries caused by mosaic equiaxed α phases in grain boundary phases.Based on the human visual attention mechanism,the attention to different visual information is different,channel attention and spatial attention are introduced into the DR-UNet encoding network,and a DRCM encoding network is reconstructed.The model pays more attention to the spatial position information of grain boundaries and can extract relatively complete grain boundary phases.Finally,in view of the scarcity of TA15 titanium alloy images,the number of images is expanded through image transformation operations such as cropping,horizontal movement,vertical movement,horizontal flipping,rotation,and scaling,and a dataset TAD of TA15 titanium alloy images is constructed.On the TAD dataset,the two algorithms proposed in this paper can extract relatively complete grain boundary αphases,which verifies the effectiveness of the above two algorithms.