Rock thin section identification is an important method in geological research. Traditional identification methods relying on polarized light microscopy has obvious shortcomings such as strong subjectivity, high cost and low efficiency. This paper proposes a rock thin plate recognition method considering multifeature fusion based on vision transformer. Vision transformer is combined with color, Histogram of Oriented Gradients and brightness features extracted manually, and the ability of global information capture is enhanced by feature stitching and transformer self-attention mechanism, aiming to overcome the limitations of traditional methods for complex rock features. The experimental results show that the accuracy of the model is 87.8%, the weighted accuracy, harmonic mean of precision and recall are 0.88, 0.87 and 0.87, respectively. Compared with the random forest model, which has higher speed and interpretability than the convolutional neural network model, the accuracy is significantly improved, which verifies the effectiveness of feature fusion and vision transformer in rock classification.