Detecting detritus from the polarized microscopic images of river sands is the first step in the tasks of sediment source analysis, tectonic evolution, and lithofacies paleogeography. Traditional detritus detection mainly relies on professionals to identify and detect manually, which is both time-consuming and labor-intensive facing large volumes of microscopic images of river sands. Currently, deep learning techniques, including Convolutional Neural Network (CNN), have achieved good performance in many visual detection tasks, and can be applied to geological tasks such as detritus detection. In this paper, we propose a novel CNN-based pesudo-siamese network for detritus detection called the SiamFuseNet. SiamFuseNet accepts both plane-polarized and cross-polarized images as input, and learns the fused feature representation to improve the detection accuracy. Besides, Both the multi-scale detection structure and the loss function are optimized to improve both the performance and robustness of SiamFuseNet. Compared to available object detection models, the experiment results show that SiamFuseNet robustly achieves greater accuracy of detritus detection without sacrificing the detection speed.

从河砂的偏振显微图像中检测碎屑是沉积物来源分析、构造演化和岩相古地理任务的第一步。传统的碎屑检测主要依靠专业人员进行人工识别和检测，面对海量的河砂显微图像，既费时又费力。目前，包括卷积神经网络（CNN）在内的深度学习技术在许多视觉检测任务中都取得了良好的性能，可应用于碎屑检测等地质任务。在本文中，我们提出了一种新的基于 CNN 的伪暹罗网络，用于碎屑检测，称为 SiamFuseNet。SiamFuseNet 接受平面偏振和交叉偏振图像作为输入，并学习融合特征表示以提高检测精度。此外，对多尺度检测结构和损失函数进行了优化，以提高 SiamFuseNet 的性能和鲁棒性。与现有的物体检测模型相比，实验结果表明 SiamFuseNet 在不牺牲检测速度的情况下稳健地实现了更高的碎屑检测精度。