Automatic lithology identification method based on efficient deep convolutional network

Due to the complexity of the environment and the variability of rocks under natural conditions, it is difficult for geologists to obtain a rapid analysis and description of rocks. To this end, this study proposes a lithology identification method that is suitable for efficient computation and maintains good accuracy. The method uses deep learning and migration learning methods to build a lithology recognition model through PyTorch and YOLOv5 frameworks, and investigates the recognition of six types of rock data. The model achieves an accuracy of 90.30% on the validation set. The method was compared with five other commonly used methods which have the fewest network parameters and can recognise 176 rock images per second on a server (equipped with a Tesla T4 GPU).