Investigation into the Feature Extraction Module and Lightweight Network for Rock Image Analysis

In scenarios such as field geological surveys, mine exploration, and engineering construction, environmental constraints often limit the deployment of large-scale computing equipment. Consequently, the variations in rock characteristics, including color, texture, and size, under different levels of complexity pose a significant challenge for outdoor rock identification. In recent years, advancements in deep learning technology have enabled real-time rock type identification, substantially enhancing the on-site decision-making efficiency of engineers, geologists, and construction personnel.This study leverages an independently constructed large-scale rock image dataset. Through the application of image cropping and augmentation techniques, the dataset was expanded to improve model robustness. A lightweight feature extraction module, termed Simple Rock Feature Extraction (SRFE), was specifically developed to address the unique characteristics of rock images. This module was integrated into a structure-optimized ResNeXt network. By refining the operational architecture and optimizing hyperparameters, the proposed model achieved efficient and accurate classification of rocks under both pristine and lens-contaminated conditions. Experimental results demonstrate that the model's highest classification accuracy reached 93.62%, rivaling the performance of medium and large neural networks. Furthermore, the model exhibited strong performance in terms of precision, recall, and F1 score. Notably, it maintains high accuracy while offering advantages such as low memory consumption, rapid inference speed, and reduced computational complexity. These features make it a reliable solution for real-time lithology identification on mobile devices and provide valuable insights for the design of lightweight network models.