WANG Mingyue, DI Yongjun^*, ZHANG Chunyu

(School of Earth Sciences and Resources, China University of Geosciences Beijing, Beijing 100083, China)

Abstract: The application of artificial intelligence in field of geosciences becomes a research hot spot in recent years. It is of great significance to the development of geosciences. One of applications of artificial intelligence is to achieve automated identification and classification of rocks or minerals using the computer vision technique. However, most researches are generally focused on the classification of rock type rather than the directly precise identification of multiple and complex minerals in the rock based on its thin section images. Although the object detection technique has been applied by many scholars to identify and classify types of rocks and minerals based on images, its application objects are mostly rock hand specimen images, and it can only be used to detect a single object in the image. In the research field of identification and classification, there is a lack of good quality algorithms and datasets for identidying and classifying minerals in the rock based on the rock thin section images. In order to solve these problems, firstly, we have collected more than 3000 images of thin sections of granite under crossed polarizer microscope, have labeled more than 10000 mineral samples on those images, have enhanced the dataset by means of data augmentation and have established a dataset with good quality and diversity. Secondly, we have proposed an improved algorithm of the RDB-Yolov5x based on the Yolov5x. In this method, the dense connection method was added in the feature extraction process, and the residual dense block (RDB) was used to have replaced the traditional residual structure. Thus, the semantic and location information details of images can be effectively preserved by using this method. Experimental results showed that this method had good generalization capability, and excellent performance for identifying small sized and fuzzy characterized mineral grains in images. By using this method, we have accurately and effectively identified five kinds of targeting minerals (quartz, biotite, muscovite, plagioclase, potassium feldspar) in the granite, with the mean average precision (mAP) up to 94.1%. Compared with the optimal comparison method, the mAP values are increased 0.5% at the IoU threshold of 50% and 1% at the threshold of 95%, respectively

Keywords: artificial intelligence identification; rocks and minerals; the RDB-Yolov5x; images taken under crossed polarizer microscope