Development and application of automatic identification methods based on deep learning for oxide scale structures of iron and steel materials.

Iron oxide structure has an important impact on the surface quality of steel, and the experimental segmentation and identification of iron oxide structure can provide help for statistical and modeling studies of iron oxide. In this paper, the images of the oxide scale structure were used to establishing a dataset, and a segmentation model was established, which was based on U-net combined with VGG16 and SENet. The results of precision statistics show that the value of PA, commonly used in the evaluation of image segmentation with the optimized U-Net was 91.56%, which was much higher than the values recognized by Deeplab V3 + , FCN-8 s, and original U-Net under the same training conditions. Based on the optimized U-Net identification network, a variety of numerical information was extracted from oxide scale images, such as the thickness range of the iron oxide scale, phase ratio of oxides, and grain size of proeutectoid Fe3O4 formed in FeO. This network can be used to calculate the structural proportion of the deformed oxide scale, further define the relationship between the proportion and its unit size, which would provide sufficient convenience and support in using images of Fe oxide structure. [ABSTRACT FROM AUTHOR]