Rock classification and identification play an important role in geologic exploration. Traditional artificial identification methods are affected by subjective factors and are cumbersome to operate. Therefore, artificial intelligence technology has been employed for the automatic identification of rock images. The rapid development of deep learning has led to several traditional classification models, but these models do not perform well on rock image data sets due to the complexity of rock features and the lack of rock data sets. To address the issue, this paper proposes a RegNet-based classification algorithm — multiscale RegNet (MS-RegNet). MS-RegNet (1) uses recursive gated convolution as the main module for feature extraction, replacing depthwise separable convolution, to enhance the global modeling capability of the model; (2) proposes a new multiscale squeeze-and-excitation attention module to enhance the feature extraction capability of the model for fine-grained features; (3) uses bidirectional feature pyramid network for feature fusion of the model, combining features of different scales to improve the classification accuracy of the model; and (4) adjusts the loss function to focal loss to reduce the negative impact of category imbalance on classification accuracy. The improved classification algorithm was tested on both private and public data sets. Experimental results show that MS-RegNet effectively focuses on the fine-grained feature differences and exhibits stronger feature extraction abilities compared with traditional classification models. The model achieved a classification accuracy and F1 score of approximately 92% on the private data set and about 94% on the public data set. This level of accuracy makes it appropriate to be used for practical applications.

You do not have access to this content, please speak to your institutional administrator if you feel you should have access.