Steel is one of the most important bulk materials these days. It is used almost everywhere - from medicine to industry. Detecting this material's defects is one of the most challenging problems for industries worldwide. This process is also manual and time-consuming. Through this study we tried to automate this process. A convolutional neural network model UNet was used for this task for more accurate segmentation with less training image data set for our model. The essence of this NN (neural network) is in step-by-step convolution of every image (encoding) and then stretching them to initial resolution, consequently getting a mask of an image with various classes on it. The foremost modification is changing an input image's size to 128x800 px resolution (original images in dataset are 256x1600 px) because of GPU memory size's limitation. Secondly, we used ResNet34 CNN (convolutional neural network) as encoder, which was pre-trained on ImageNet1000 dataset with modified output layer - it shows 4 layers instead of 34. After running tests of this model, we obtained 92.7% accuracy using images of hot-rolled steel sheets.

Nowadays, many different tools can be used to classify images, each of which is aimed at solving a certain range of tasks. This article provides a brief overview of libraries and technologies for image classification. The architecture of a simple convolutional neural network for image classification is built. Image recognition experiments have been conducted with popular neural networks such as VGG 16 and ResNet 50. Both neural networks have shown good results. However, ResNet 50 overfitted due to the fact that the dataset contained the same type of images for training, since this neural network has more layers that allow reading the attributes of objects in the images. A comparative analysis of image recognition specially prepared for this experiment was carried out with the trained models.

This article presents solutions to the person's fatigue recognition problem by the face's image analysis based on convolutional neural networks. In the present paper, existing algorithms were considered. A new model's architecture was proposed and implemented. Resultant metrics of the model were evaluated.