Classification Of Rock Slices Based On SKNet

When classifying rocks in the field of petroleum mining and beneficiation, factors such as rock type, particle size, and interstitial spacing make the rock classification task complicated and low in accuracy. When manual classification is used, the workload is large and it is easily affected by subjective factors. This paper proposes a rock multi-classification method based on Selective Kernel Networks (SKNet). The method first builds a Residual Network (ResNet) model to classify rock slice images, and the accuracy of this method reaches 89.5%. At the same time the Gradient-weighted Class Activation Map (Grad-Cam) is used to visualize the model and increase the interpretability of the neural network. Finally, SKNet is constructed on the existing ResNet101 model to improve the accuracy of classification. After many experiments, it is finally accurate the rate reached 95.3%. In this experiment, the central-northern rock slice images of the Lhasa terrane are used to complete the experiments and tests. The experimental results show that using SKNet to classify rock slices can achieve faster and more accurate classification results.