A survey of ore image processing based on deep learning

Ore is an essential industrial raw material and strategic resource that plays an important role in China’s economic construction. The smart mine aims to build an unmanned, efficient, intelligent, and remote factory to improve quality, reduce cost, save energy, and increase the efficiency of mineral resource extraction. Ore image processing technology can automatically and efficiently complete a series of difficult and repetitive tasks, which constitutes an important part of smart mine construction. However, open-air operation modes, high-dust environments, and ore diversity have brought great challenges to ore image processing. Benefiting from its strong automatic feature extraction ability, deep learning can deeply perceive a complex environment, which enables it to play an important role in the ore image processing field and help traditional mining companies transform into efficient, green, and intelligent enterprises. This paper focuses on two production stages, including ore prospecting and belt transportation. We systematically summarize the main applications of deep learning in ore image processing, including ore classification, particle size analysis, and foreign material recognition, sort out the corresponding algorithms, and analyze their advantages and disadvantages. Specifically, according to the number of ores in an image, ore classification is divided into single-object and multi-object classifications. Single-object classification is mostly addressed by image classification networks, while multi-object classification is mostly accomplished by object detection and semantic segmentation networks. Single-object classification plays an important role in geological prospecting. Particle size refers to the size information of ores in an image. Generally, it can be divided into three modes: particle size statistics, particle size classification, and large block detection. Among these modes, the first and the third are mainly used in actual industrial production. Particle size statistics are determined mostly using semantic segmentation networks and can provide a reference for the control of crushers and conveyor belts. Large block detection is performed mostly by adopting object detection networks and can identify the oversized ore on an ore feeding belt and prevent material blockage accidents in the transfer buffer bin between the ore feeding belt and the ore receiving belt. Foreign material recognition detects harmful objects mixed in the ores on the belt to ensure product quality and prevent the belt from tearing. Object detection technology is often used to complete the task of foreign material recognition.