AI Network Reads Rotating Polarizer Views to Trace Thin Mineral Grain Edges

Grain boundaries in cross-polarized petrographic thin sections do not look the same from every angle. As the polarizer rotates, some boundaries brighten into sharp edges while others fade into extinction, which makes automatic boundary extraction highly sensitive to whichever single view an algorithm happens to analyze. A new study addresses this instability with a deep learning model that deliberately chooses among views rather than committing to one.

The method, called the View-Selective Thin-section Anisotropic Edge network, or VISTA-Edge, was developed by Lanlan Zhang, Chenyang Zhu, Jierui Wang, Yunxin Xie and Fang Wang, and is described in Earth Science Informatics. Instead of treating the registered polarized observations as redundant copies, the network applies view-selective attention that weights each input at every pixel according to its learned agreement with a mean-reference descriptor, producing a fused RGB image that preserves the boundary information scattered across the rotation sequence.

Architecturally, VISTA-Edge pairs an EfficientNetV2 encoder with a directional strip decoder that aggregates multi-scale features along horizontal and vertical neighborhoods, a design suited to the elongated, anisotropic character of mineral grain edges. High-resolution encoder guidance then refines the boundary response, keeping fine edge detail that coarse decoding stages would otherwise blur. Training relies on a composite loss that combines balanced classification, overlap, centerline alignment, and suppression of responses away from true edges.

At inference, the predicted probability maps pass through hysteresis thresholding and skeletonization, converting soft edge responses into clean boundaries a single pixel wide. The model was evaluated on the patch-level split of the CPPID dataset using seven polarized observations together with their maximum-response composite. VISTA-Edge reached BF1@1, AP, and ODS scores of 87.87, 67.34, and 64.33, exceeding MLEEP, the strongest external baseline in the comparison, by 1.20, 3.65, and 2.29 percentage points respectively.

Component ablations reported in the abstract point to edge-guided upsampling as the largest contributor to the measured gains, suggesting that maintaining spatial resolution at the decoding stage matters as much as selecting the right views. The reported figures come from the published abstract; the full experimental details sit behind the journal’s subscription access.

Subject of Research: A view-selective deep learning network for extracting grain boundaries in petrographic thin sections.

Article Title: Multi-view extinction-guided thin-edge segmentation for petrographic thin sections

Article References: Zhang, L., Zhu, C., Wang, J., Xie, Y., & Wang, F. (2026). Multi-view extinction-guided thin-edge segmentation for petrographic thin sections. Earth Science Informatics, 19(12), Article 232. https://doi.org/10.1007/s12145-026-02286-5

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02286-5

Keywords: petrographic thin sections, grain-boundary segmentation, cross-polarized light, deep learning, view-selective attention, edge detection, EfficientNetV2, CPPID dataset, Earth Science Informatics, computer vision, Multi-view, extinction-guided