An Improved Mineral Image Recognition Method Based on Deep Learning

Identification of rock minerals is one of the fundamental procedures of geology and mineralogy. Computer vision technologies and the theory of deep learning (DL) make intelligent rock mineral identification possible. The polarizing microscope pictures of iron ore as the data source, a composite dataset consisting of transmitted light images and reflected light images was developed in this investigation. Using the Deeplabv3+ network, a targeted mineral identification network model was created based on the DL theory. This model can effectively and automatically extract the deep feature information of ore mineral images under a polarizing microscope, as well as achieve intelligent identification and classification of transparent minerals and non-transparent minerals. The model was then improved by freezing training, enlarging the receptive area, and utilizing FC-CRF. The outcome demonstrated the outstanding performance. The total mineral recognition accuracy reached 97.56%, and the recognition accuracy of certain minerals was up to 99%. The identification result obtained by the improved mineral identification model accurately depicts the mineral species information of the microscope photographs, providing a convenient and trustworthy data source for the development of intelligent mineralogy.