Rock image lithology recognition method based on lightweight convolutional neural network

ObjectiveLithology identification is a crucial step in the process of oil and gas detection and exploration, providing important guidance for exploration positioning, reservoir evaluation, and the establishment of reservoir models. However, traditional manual lithology identification methods are time-consuming and labor-intensive. Although classical deep learning models achieve high identification accuracy, they often have a large number of parameters. To enhance model accuracy while reducing the number of parameters, the aim of this research is to make the model suitable for real-time lithology identification. MethodsThis paper first collected a dataset of 3016 rock images consisting of eight types of rocks, including dolomite and sandstone. Based on the lightweight convolutional neural network ShuffleNetV2, the paper proposes a Rock-ShuffleNetV2 lithology identification model (RSHFNet model). The model incorporates the convolutional block attention module (CBAM) and multi-scale feature fusion module (MSF) into the basic network to enhance feature extraction capabilities and improve identification performance. Additionally, the number of stacked ShuffleNetV2 units is optimized to reduce the model's parameters. ResultsThe experimental results show that the RSHFNet model achieved an accuracy of 87.21%, which is a 4.98% improvement over the baseline model. Furthermore, the model's parameters and floating-point operations were reduced to 869702, 0.93×108, respectively, representing 67% of the model's parameters and 63% of the floating-point operations of the baseline model. This reduction significantly decreases the model's size. Additionally, the RSHFNet model demonstrates superior overall performance compared to existing convolutional neural networks. ConclusionThe proposed RSHFNet lithology identification model offers high recognition accuracy and strong generalization capabilities while being more lightweight, providing a new approach for real-time lithology identification in the field.