Improving Rock Type Identification Through Advanced Deep Learning-Based Segmentation Models: A Comparative Study

Abstract

The accurate identification of rock types is crucial for understanding geological structures and planning mining activities. Therefore, the precise labeling of rock types is a fundamental requirement for researchers and industry experts in these fields. This study aims to identify rock types by segmenting thin-section rock images using advanced deep learning models. Commonly used models such as DeepLabv3+, SegFormer, ConvNext, and Mask2Former were evaluated to compare the performance of different segmentation models. Additionally, an improved version of Mask2Former was analyzed to enhance its performance. Post-segmentation enhancements with SLIC super pixel and morphological operations further improved boundary delineation. An improved version of Mask2Former achieved a validation accuracy of 91.26% and a mean Intersection over Union (mIoU) of 82.59%, representing an improvement of more than 5% over the base Mask2Former. Another significant aspect of the study is the detailed analysis of other models using the relevant dataset. The Quartz Feldspar Plagioclase (QAP) diagram was utilized for mineral identification, achieving a mineral recognition accuracy of 85.7%. These results indicate the robustness of the proposed approach, which exceeds the accuracy and mIoU of comparable methods reported in the literature. This study significantly enhances the effectiveness of image processing techniques for rock type identification. Furthermore, the detailed comparison of different models provides valuable guidance for researchers in selecting the most suitable model.