Abstract: The accurate identification of roads in open-pit mines is very important for the digitization and unmanned intelligent construction of mining areas. However,the existing machine vision methods are difficult to accurately identify shadow occlusion and gravel accumulation roads. A VMA-UNet network for road recognition of UAV images in mining areas is proposed,which replaces the original U-Net backbone network with VGG16,and combines polarization multi-scale feature self-attention (PMFS) and channel space parallel attention (CSPA) mechanisms. The UAV was used to collect the road images of the mining area,which were divided into three types:unstructured,shaded and gravel accumulation roads. After CLAHE processing,they were labeled with Labelme to make multi-type road data sets and transfer learning was used to pre-train the model. Comparative experiments show that the CLAHE operation can significantly highlight the characteristics of different types of roads in the mining area,and the transfer learning makes the network converge faster and better adapt to the mining area road data in different scenarios. VMA-UNet performs well in various mining area road datasets. The accuracy,recall and average intersection-union ratio on the total dataset are 94. 8%,91. 29% and 80. 58%,respectively. The indexes are better than D-LinkNet,DeepLabV3+ and U-Net network models,which can accurately identify the edge of shadow occlusion and gravel accumulation roads,and effectively improve the recognition accuracy of multi-type roads in mining areas. 

Key words: UAV,deep learning,multi-type roads for open-pit mines,transfer learning,road recognition,attention mechanism