Research on Lightweight Lithology Intelligent Recognition System Incorporating Attention Mechanism

How to achieve high-precision detection and real-time deployment of the lithology intelligent identification system has significant engineering implications in geotechnical, geological, water conservation, mining disciplines. In this study, a lightweight model is proposed overcome problem. The MobileNetV2 utilized as basic backbone network decrease operation parameters. Furthermore, channel attention spatial methods are incorporated into improve network’s extraction complicated abstract petrographic elements. addition, based on findings training, computing power performance, test results, Grad-CAM interpretability analysis comparison tests with Resnet101, InceptionV3, models. training accuracy 98.59 percent, duration 76 min, trained just 6.38 megabytes size. precision (P), recall (R), harmonic mean (FI-score) were, respectively, 89.62%, 91.38%, 90.42%. Compared three competing models, presented work strikes better balance between recognition speed, it gives greater consideration rock feature area. Wider more uniform, strong anti-interference capability, improved robustness generalization performance model, which can be deployed client or edge devices some promotion value.