Improving the accuracy of semantic segmentation of carbides in the microstructure of composite coatings by the neural network

The volume fraction of primary carbides is one of the indicators that affects the wear resistance of composite coatings. To calculate the volume fraction of carbides, a multiclass (carbides, pores, the rest of the microstructure and areas not related to the microstructure) semantic segmentation of images with the structure of composite coatings by the neural networks based on DeepLab-v3 was carried out. Various approaches to the neural networks training are considered: resizing all images of the training set to some fixed size using interpolations, slicing images into fixed -size fragments. In addition, we varied in a wide range of mini -batch size and learning rate. It is shown that the best result of the network performance according to the mIoU metric was obtained for the network trained with a mini -batch size of 320 and a learning rate of 1 x 10-3 using a method in which training and validation were carried out on statically augmented fragments of original images, and the processing of test set was carried out by double analysis, consisting of averaging the results for fragments of the original and resized to half images. The results obtained were characterized by a smaller calculation error compared to the calculation according to ASTM E 562-02.