Microscopic Image Segmentation of Pure Iron Grains Based on Deep Learning
Metallographic analysis is an important means of 3D construction of grain structure.To realize the calculation,mea-surement and visualization of the 3D spatial morphology of grains accurately,efficiently and quickly,it is necessary to extract the grain boundaries in the metallographic structure accurately.Traditional machine learning and deep learning models are extremely susceptible to noise interference such as grain boundary blurring,disappearance,and scratches during the extraction process,re-sulting in the inability to accurately extract grain boundaries.This paper proposes a convolutional neural network model GAU-Net that fused U-net and Conv Gate Recurrent Unit(convGRU).It aimes to solve the problem that the traditional convolutional neural network model cannot obtain the spatial trajectory information of the related pictures.The GAU-Net model ensures the input of the original image and correlates the high-level features of the temporal and spatial domains of the previous image by dual image map-ping.It uses a feedback mechanism that imitates human brain thinking.When the second round of network extracted features,the high-level features of the first round would be fused,and different types of feature fusion would be performed based on the size and dimension of the feature map.Model lightweighting is achieved while avoiding image feature loss.The findings show that,for the pure iron grain slice dataset,compared with other classical model algorithms,the method in the paper can accurately segment the grain boundaries in the complex environment.
grader extractiondual mode picture mappingspatial and space characteristicsmulti-dimensional feature fu-sionmodel lightweight
国家科技期刊平台
NSTL
万方数据
