Determination of the volume fraction of primary carbides in the microstructure of composite coatings using semantic segmentation

In the process of formation of composite coatings, partial dissolution of hardening particles (most often carbides) in the matrix is possible; therefore, in some cases, the material creation mode is chosen taking into account the volume fraction of primary carbides not dissolved during coating deposition. The methods currently widely used for calculating the volume fraction of carbides in the structure of composite coatings (manual point method and programs implementing classical computer vision methods) have limitations in terms of the possibility of automation. It is expected that performing semantic segmentation using convolutional neural networks will improve both the performance of the process and the accuracy of carbide detection. In the work, multiclass semantic segmentation was carried out including the classification on the image of pores and areas that are not a microstructure. The authors used two neural networks based on DeepLab-v3 trained with different loss functions (IoU Loss and Dice Loss). The initial data were images of various sizes from electron and optical microscopes, with spherical and angular carbides darker and lighter than the matrix, in some cases with pores and areas not related to the microstructure. The paper presents mask images consisting of four classes, created manually and by two trained neural networks. The study shows that the networks recognize pores, areas not related to the microstructure, and perfectly segment spherical carbides in images, regardless of their color relative to the matrix and the presence of pores in the structure. The authors compared the proportion of carbides in the microstructure of coatings determined by two neural networks and a manual point method.