A lightweight model hyperparameters searching method for fast, accurate and on-site lithology identification

The efficiency of tunnel excavation, rock strength, stability of surrounding rock, and underground engineering disasters are closely related to lithology. Accurately identifying lithology is a necessary prerequisite for intelligent, safe, and efficient tunnel construction. The design of conventional recognition models heavily relies on experience and extensive calculations. To develop a model suitable for deployment on construction sites and capable of accurate lithology identification, a fast search method for lithology identification models is proposed. This method integrates geological knowledge, apparent feature extraction techniques, and search algorithms. An efficient feature extraction super network using multi-scale geological features of rock surface is constructed, a model evaluation method that comprehensively considers accuracy and latency is developed, and differential evolution algorithm is used to search for the optimal model parameters. Experiments demonstrate that the proposed method enables the model to evolve faster and more accurately, and eventually a model (LithoNet) suitable for lithological classification is found. It only takes 2.10 ms to infer an image of 224 × 224, which is 57.25% faster than MobileNet v3 and 62.83% faster than ShuffleNet V2. The F1-score of LithoNet is 0.9874, surpassing classical models such as EfficientNetV2-S. LithoNet can be easily deployed on portable devices, effectively promoting the intelligence and accuracy of lithology identification at engineering sites.