The article presents an embodiment of an artificial neural network for recognizing defects in images of steel sheets. Several stages of solving the problem are described: the choice of a development environment, a programming language, and libraries necessary for the implementation; features of data analysis, graphing, histograms, finding dependencies; the selection of a suitable neural network, the choice of neural network architecture, the selection of an algorithm for assessing quality and accuracy; neural network spelling; training and checking accuracy and quality, checking for overfitting (retraining). As development tools, Python language, PyTorch library, Jupyter development environment, convolutional neural network architecture — Unet are proposed. Features of the analysis of input images of steel sheets, features of the implementation of the neural network itself are described. The function of binary cross entropy was chosen as a criterion for assessing accuracy, since it seeks to bring the distribution of the network forecast to the target, fine not only for erroneous predictions, but also for uncertain ones. For additional evaluation, the DICE method was also used. The accuracy of the resulting model is 84 %. The proposed solution can become part of a hardware-software system for automating the recognition of defects on metal sheets.
Separating and denoising seismic signals with dual-path recurrent neural network architecture
