Quick recognition of rock images for mobile applications

Abstract

Rock lithology identification is an important aspect of field geological surveys. However, traditional identification methods cannot obtain real-time effective feedback, a limitation that is ineffective for further implementation of field geological surveys. To obtain rock lithology information quickly in field geological surveys, an automatic identification method of rock lithology applicable to field offline conditions was proposed in this study. Based on MobileNets, a lightweight deep neural network with depthwise separable convolutions, and the transfer learning method, the proposed method was employed to establish a lithology recognition model for rock images. Its applications on mobile devices were verified, and the fast and accurate lithology identification of rock images was realized under the field offline conditions. Results demonstrate that the constructed model achieves 95.02% identification accuracy on the validation dataset and 93.45% identification accuracy on the test dataset of mobile devices. The average recognition time of each image is 1186 ms, and images which have result confidence higher than 95% respectively account for 91% of the test dataset. The mode size is 17.3 MB. These findings indicate that the model has high identification accuracy, short identification time, and reliable identification results. The proposed method provides good references for the establishment of a geological survey intelligent space and has promising prospects for application in field geological surveys.