The application of deep neural networks for the classification of the microstructure of cast iron products

The possibilities of using artificial neural networks in the tasks of recognizing the microstructure of cast irons have been investigated. Convolutional neural networks work on the basis of filters that are concerned with recognizing certain characteristics of an image (for example, straight lines). A filter is a collection of kernels; sometimes a single kernel is used in a filter. A kernel is a common matrix of numbers called weights that are "trained" in order to search for certain characteristics in images. The filter moves along the image and determines if some desired characteristic is present in a particular part of it. To obtain an answer of this kind, a convolution operation is performed, which is the sum of the products of the filter elements and the matrix of input signals. If some desired characteristic is present in the image fragment, the convolution operation will produce a number with a relatively large value at the output. If the characteristic is absent, the output number will be small. Note that the number of filter channels must match the number of channels in the original image; only then will the convolution operation produce the desired effect. For example, if the original image consists of three channels, the filter must also have three channels. The structure (architecture) of the convolutional neural network was proposed and its ability to recognize the presence of structural components of the microstructure of cast iron was confirmed. The training of the network was performed on images of microslices of various types of cast iron, and positive results were achieved with an accuracy close to 85%. The results of the work indicate the prospects of using convolutional neural networks in the tasks of recognizing and classifying the microstructure of cast iron.