基于深度学习的岩石铸体薄片图像孔隙自动提取

岩石铸体薄片图像中孔隙区域的准确提取是分析评估工作的前提。但目前传统的孔隙提取方法主要是通过颜色特征进行阈值分割，精度较低，需加入大量的人工交互操作才能提高精度。因此提出一种新的基于深度学习的孔隙区域自动提取算法，该算法基于U-net搭建网络基本框架。首先，网络在编码阶段加入残差块来提升网络的深度。其次，针对残差块进行优化并引入空洞卷积，提取更全局、语义层次更深的特征。最后，在解码阶段加入网络模块间的短连接，提出新的融合特征方法，更好地将浅层特征与深层语义特征相结合，得到更加精细的孔隙区域。实验结果表明，该方法优于传统的孔隙提取方法，具有更高的分割精度且无需人工操作，与主流卷积神经网络相比也具有更高的精度和平均交并比。

Automatic extraction of pores in thin slice images of rock castings based on deep learning

Accurate extraction of the pore region in the image of rock casting thin section is the premise of analysis and evaluation. However, the traditional methods of pore extraction mainly use color features for threshold segmentation, which requires a lot of manual interaction to improve the accuracy. Therefore, a new deep learning-based automatic pore region extraction algorithm is proposed, which is based on U-net to build a basic network framework. First, the residual blocks is added to the network during the encoding phase to improve the depth of the network. Secondly, the residual block is optimized and the hole convolution is introduced to extract more global and deeper semantic features. Finally, a short connection between network modules is added at the decoding stage, and a new fused feature method is proposed to better combine shallow features with deep semantic features to obtain a finer pore area. The experimental results show that the method is superior to the traditional pore extraction method, has higher segmentation accuracy and does not require manual operation, and has higher accuracy and average intersection ratio than mainstream convolutional neural networks.