This study is devoted to the development of methods for the automatic assessment of the granulometric composition of ore after blasting based on data obtained from unmanned aerial vehicles (UAVs). Determining the size of ore fragments is an important step, since the effectiveness of subsequent crushing processes depends on its accuracy. Traditional methods of analysis use manual work, which requires considerable labor and is subject to subjective factors. The study examines modern machine learning methods and neural network architectures, such as Feature Pyramid Network (FPN), EfficientNet and SE ResNet, which can automatically and accurately segment images. As a result of the experiments, it was found that the VPN network with a pre-trained EfficientNet B2 base showed the highest IoU accuracy among the models.

Keywords: granulometric composition, FPN, Efficient Net, SEResNet

The drilling and blasting method is currently the most widely used for mining rocks. An indicator of the high–quality drilling and blasting operations is the uniform granulometric composition of the exploded rock mass - the percentage of oversized ore pieces should be minimal. The percentage of oversized and its increase have a significant impact on the technical processes of transporting rock mass, leading to an increase in the costs of loading and transportation operations and secondary crushing of oversized ore masses. The paper describes the results of a study of methods for determining the granulometric composition of drilling and blasting operations using neural networks of segmentation Unet and FPN. Images taken from UAVs are used for analysis. A method of classifying ore by size has also been developed, which ensures the accuracy of the proportion of correct answers of more than 0.91. The expected result of the introduction of the system for automatic determination of the granulometric composition of drilling and blasting operations is the possibility of more accurate control over the quality of their performance.

Keywords: granulometric composition, Unet, FPN, classification, segmentation

This article describes the results of research on the possibility of detecting sunflower plants from photographs taken by a UAV. The solution to this problem will allow automated control of an important agricultural parameter - seedling density. The problem is complicated by the limited amount of training sample and "disturbances" associated with field weeding. As the result we obtain that the YOLOv5m neural network is capable on a sample of 122 pictures to qualitatively detect plants with an error of 0.077% of training. Artificially increasing the sample to 363 pictures reduces the learning error to 0.063%. Disturbances reduce the detection efficiency of sunflower plants in the test images. It is possible to increase the detection efficiency either by adding original images to the training sample or by artificially enlarging the sample.

Keywords: detection, YOLOv5, sunflower, seedling density, neural network