Abstract:In rock image lithology identification, deep learning frequently encounters challenges such as background interference and the difficulty in distinguishing similar lithology categories. To address these issues, this paper introduces a two-stage parallel rock image classification method based on an enhanced YOLOv11 network. In the first stage, the YOLOv11 segmentation model is employed to eliminate background noise from rock images and isolate the target rock objects. In the second stage, rock objects obtained through sliding window clipping are classified in parallel using the improved YOLOv11n classification model, with threshold processing applied to the prediction results to enhance the accuracy of classifying similar lithology categories. To improve the model"s ability to capture detailed features of rock images, this study incorporates a super-resolution SAFMN reconstruction module into the YOLOv11n classification model, adapting it to fit the YOLO network architecture. Additionally, the loss function of the YOLO network has been optimized by adopting the GHM-Loss strategy, which balances the contribution of different sample types during model training, thereby enhancing both the efficiency and reliability of the training process. This paper constructs datasets for rock image segmentation and classification and conducts relevant experiments on these datasets. The comparative experimental results demonstrate that the YOLOv11 segmentation model outperforms other models in rock image segmentation tasks. Compared to the baseline model, the enhanced YOLOv11n classification model achieves improvements of 12% in precision (P), 1.7% in recall rate (R), and 0.4% in accuracy rate (A). In the final method comparison experiment, the proposed two-stage parallel rock image classification method demonstrates an average improvement of 22.5% in accuracy, validating its superiority and effectiveness in rock image classification tasks.