An innovative adaptive mineral segmentation method via augmentation and fusion of single and orthogonal polarized images: AMS-p/xpl

Abstract

An adaptive mineral segmentation method is proposed to address the challenges of traditional thin section identification, which is typically expert-dependent, highly subjective, and time-consuming. The method is based on the augmentation and fusion of single polarized (PPL) and orthogonal polarized (XPL) images using an enhanced Deeplabv3-based segmentation model. The Depth Separable Convolution (DSC) is introduced to strengthen edge and texture features from PPL images, and a ColorBoost module is designed to enhance color information from XPL images. An adaptive feature fusion mechanism is employed to integrate complementary polarized features and dynamically adjust their contribution weights. The results demonstrate that the proposed model achieved the highest segmentation performance on the test set, with a mean intersection over union (mIoU) of 89.0 % and an accuracy of 96.7 %. Compared to widely used semantic segmentation networks such as FCN, it demonstrates a notable improvement in mIoU, with a maximum gain of 32.9 %. Additionally, through the integration of feature augmentation and an adaptive fusion mechanism, the model outperforms the baseline DeepLabv3 by 5 % in mIoU. The proposed method provides a more efficient and automated solution for thin section mineral identification, reducing reliance on expert knowledge and improving applicability in practical and non-specialist settings.