Multi-channel attention transformer for rock thin-section image segmentation

Accurate rock thin-section image segmentation can help to analyze the chemical composition, particle size distribution, pore structure and cement composition. However, precise instance segmentation is currently challenging due to the issues of small sample size, lack of integration of sequence images with different lighting angles and low representation learning capability. To address the aforementioned challenges, this paper introduces a groundbreaking Multi-Channel Attention Transformer (MCAT) approach for rock thin-section image segmentation. At first, the copy paste method is applied for data augmentation to overcome the small sample issue. Secondly, a novel multi-channel attention module is developed to integrate the correlation between the image sequence derived from different lighting angles. Finally, the powerful Transformer module is employed to enhance feature learning. The experiments conducted on the real rock thin image dataset validate the superiority of the proposed MCAT approach over the existing methods.