Применение глубокого обучения для сегментации камней на конвейерах и складах горнодобывающих предприятий

Abstract

In the last decade, the introduction of artificial intelligence methods in industry has been accelerating. The development of deep learning algorithms and the emergence of the ability to store and process large amounts of information make it possible to quickly and efficiently automate tasks that previously could only be solved by people – employees of enterprises, and the results obtained not only correspond to human cognitive abilities, but often surpass them. An interesting example of a routine task that can be automated using computer vision methods is the task of segmenting stones on conveyors and warehouses of mining enterprises to ensure quality control of raw materials and finished products. The purpose of this work is to develop an algorithm for segmenting stones on conveyors and warehouses. To achieve this goal, a brief historical review of approaches to solving the described problem was carried out, and a study was made of the application of the Mask R-­CNN architecture to solving the problem of stone segmentation. The training dataset included 1000 augmented images from 100 crushed stone photos taken on a mining conveyor belt. The results obtained in the IoU metric exceeded 83 %, and in the Accuracy metric – 89 %, which provides high-quality automatic continuous visual quality control of raw materials or finished products. The resulting segmentation maps can serve as a good basis for determining granulometric characteristics, quality categories that are important in the mining industry, timely detecting flakiness on conveyors and segregation in finished product warehouses in real time.