Year: 2024
Type: article
Abstract: Drill core image recognition is an extremely critical aspect of geological exploration.However, the research on drill core image recognition faces challenges such as subtle discrepancies between different lithological categories, complex backgrounds, unclear recognition subjects, and lack of datasets.Currently, the existing recognition methods for geological images struggle to accurately classify drill core images.To address these issues, this paper con-structs a drill core image dataset called Drill Core Dataset (DCD), and proposes an image recognition model called MSL-ResNet50, which combines self-calibration residual module and three-dimensional attention.By introducing a self-calibration module with a larger receptive field, the model captures higher-level semantic information and enhances the network's feature transformation capability.Furthermore, it integrates the SimAM attention mechanism to mitigate interference in context information, giving higher attention to key areas.Additionally, the MAM pooling strategy is utilized to alleviate the loss of image features and maintain consistency in feature distribution of feature maps before and after pooling.Extensive ablation and validation experiments were conducted on the DCD dataset.The experimental results demonstrate that the proposed model, MSL-ResNet50, significantly outperforms the related benchmark models in drill core image recognition.Specifically, it achieves a recognition accuracy of 99.3%, an improvement of 4.1% over the original model, validating the effectiveness of the proposed model for core image recognition tasks.
Authors: Yong Pu, Zhihao Shao, Ziyuan Xu, Chuanhu Xiong, Yonghua Chen
Institutions: Guangzhou Metro Design & Research Institute, Hohai University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4406357800