Computer Vision and Deep Machine Leaming in Image Analysis of Hard Alloys

Introduction: The is related to the microstructure properties of tungsten-cobalt carbide-cobalt carbide (WC-Co) alloy from the perspective of neural network feature space. Such features are called hidden features and they are often obtained using neural network autoencoders. The peculiarity of these features is that they store much more information than the human eye can see in an image. Our work is aimed to analyze the feature space and create methodologies to classify alloys based on the data. Methods: We applied methods from deep machine learning, in particular we applied a modification of the classical variational autoencoder (VAE) called VQ-VAE-2. Neural networks with VQ-VAE-2 architecture show high performance in the hidden feature extraction task and at the same time are not fully explored. We also applied classical methods from linear algebra and statistics. Results: Analysis of the application of VQ-VAE-2 showed that top and bottom features represent different information about the input data. In particular, top is responsible for color transfer, while bottom is responsible for image structure. Experiments with dimensionality reduction have shown that the Umap algorithm is able to efficiently form structures in the two-dimensional VQ-VAE-2 feature space. At certain parameters we obtain structures with good clusters, by which we can classify the alloy classes. The latent feature matrix has a random spectrum of values, and also at SVD decomposition has a random distribution of singular numbers. No correlations have been found. Methods of generating new images using VQ-VAE-2, resampling and Umap inverse transform were considered. All of them did not show high generation quality, but gave a good understanding about image generation. Discussion: Our results can be useful to material scientists and engineers. We hope to provide more information about the behavior and properties of the alloy, which in turn will help experts to work with it more effectively.