Rock Thin Section Image Identification Based on Convolutional Neural Networks of Adaptive and Second-Order Pooling Methods

Abstract

In order to enhance the ability to represent rock feature information and finally improve the rock identification performance of convolution neural networks (CNN), a new pooling mode was proposed in this paper. According to whether the pooling object was the last convolution layer, it divided pooling layers into the sampling pooling layer and the classification pooling layer. The adaptive pooling method was used in the sampling pooling layer. The pooling kernels adaptively adjusted were designed for each feature map. The second-order pooling method was used by the classification pooling layer. The second-order feature information based on outer products was extracted from the feature pair. The changing process of the two methods in forward and back propagation was deduced. Then, they were embedded into CNN to build a rock thin section image identification model (ASOPCNN). The experiment was conducted on the image set containing 5998 rock thin section images of six rock types. The CNN models using max pooling, average pooling and stochastic pooling were set for comparison. In the results, the ASOPCNN has the highest identification accuracy of 89.08% on the test set. Its indexes are superior to the other three models in precision, recall, F1 score and AUC values. The results reveal that the adaptive and second-order pooling methods are more suitable for CNN model, and CNN based on them could be a reliable model for rock identification.