Research On Feature Recognition Of Rare Earth Mining Based On Improved Tiny-YOLOv3 Model

Rare earth is an important strategic resource in China,which is of great significance to national economic and strategic development.However,due to the increasing demand of rare earth resources,its mining intensity is also greater,at the same time,illegal exploitation of rare earths resources,serious destruction of natural ecological environment phenomenon is more serious.With the rapid development of remote sensing investigation technology and artificial intelligence,automatic recognition of mining characteristics in rare earth mining areas has become possible,which is of great significance to the real-time monitoring of rare earth mines.Based on the project of "Extraction of Feature Information of Non-Oil and Gas Mines in Wuyishan Metallogenic Belt of Western Hunan in 2019," Using BJ-2 satellite image data,Dingnan County and Longnan County in southern Jiangxi Province as study area,the mining characteristics of rare earth mining area are identified and detected,including circular sedimentation tank and square high level tank.In order to ensure a higher detection speed and precision,the deep learning target detection algorithm is deeply understood.Based on the Tiny-YOLOv3 model in single-stage target detection,we improve its network structure to achieve our goal.The main research results are as follows:(1)Tiny-YOLOv3 is selected as the baseline of this improved algorithm,and two kinds of prediction layers are selected.Based on YOLOv3 and Tiny-YOLOv3,the performance of the detection results is compared.The results show that although the recognition accuracy of YOLOv3 is higher than that of Tiny-YOLOv3,However,its processing efficiency is only one quarter of that of Tiny-YOLOv3,and the processing process is complicated.Therefore,we choose to improve the model based on Tiny-YOLOv3.At the same time,by comparing the performance effect of TinyYOLOv3 in 2 prediction layers and 3 prediction layers,it is found that The detection effect is better when the prediction layer structure of 13 × 13 and 26 × 26 is preserved.(2)Based on the Tiny-YOLOv3 network model,by embedding RFB-SE network,adding convolution layer and shortcut,introducing spatial pyramid pooling network and replacing three convolution layers with Coord Conv layer,the original network structure is optimized.(3)Improved Tiny-YOLOv3 model improves target detection accuracy.Compared with the original Tiny-YOLOv3 network and the improved Tiny-YOLOv3 network,the results show that the detection accuracy of the modified Tiny-YOLOv3 has been greatly improved.The m AP value is 94.6%,which is nearly 3% higher than the original network model,and the processing speed is still high,the FPS value is 87.5.(4)The improved network model is better than other classical target recognition detection algorithms.In this paper,the improved Tiny-YOLOv3 model and Faster R-CNN,SSD300 and Retina Net are compared.The results show that the improved model has higher detection accuracy than SSD300 and Retina Net,and is close to Faster R-CNN,and its processing speed is much faster than the other three models.The improved model has more advantages on the whole from the comprehensive evaluation indexes of detection precision and processing speed.