A Research On Automatic Segmentation Methods For Rock Thin Section Images

Analysis of rock thin sections is an important tool for understanding geological composition,as well as oil and gas exploration and development.Image segmentation,which divides an image into multiple sub-regions of mineral particles,is the first step in the identification process.The results are crucial for comprehending the physical properties of geological layers,creating oil and gas development plans,and calculating reserves.Traditional manual rock thin section segmentation has limitations,including high expertise,time and effort,a lack of accuracy,and subjective factors,while existing automated solutions are expensive and ineffective for minerals with similar chemical properties.Currently,with the popularity of deep learning methods in the field of computer vision,deep neural networks have made it possible to achieve end-to-end intelligent recognition and segmentation of rock thin-section image particles due to their stronger feature extraction capability and good performance.Therefore,this thesis primarily investigates the automatic segmentation technique of rock thin-section images under polarized light microscopy,and the main works are as follows:(1)An extinction consistency perception network is proposed for the edge detection of mineral particles in rock thin-section images under cross-polarized light.The model extracts the extinction characteristics of mineral particles using the proposed multi-angle extinction consistency module,and then the multi-angle polarized image sequences are fused into edge-enhanced output features.Meanwhile,a multi-scale edge-aware network is built utilizing modified Efficient Net V2 as the encoder and a bidirectional feature flow decoder to achieve different levels of feature representation.Meanwhile,a distance map penalized composite loss function is introduced to direct the model’s attention to edges.Experimental findings show that the proposed model is effective,which scores of 0.940 ODS and 0.941 OIS and were about 15% higher than seven classic edge detection models.(2)The Rock Trans Net network is proposed for mineral particle instance segmentation of rock thin section images under cross-polarized light.The Mselect copy&paste data augmentation algorithm based on mineral particle instance filtering is proposed for data augmentation,which increases the diversity of mineral particle types and sizes,alleviates the model overfitting problem,and allows it to predict mineral particles that are not provided in the ground truth.Meanwhile,a feature shift algorithm and a lightweight ECA channel attention module are introduced to improve the original extinction consistency image fusion module,resulting in a two-thirds reduction in training time.Finally,the Mineral Ins Seg Net with Swin-transformer as the backbone network and Mask R-CNN as the instance segmentation head is proposed to achieve the final instance segmentation of mineral particles.The model accomplishes first place among five relevant instance segmentation metrics when evaluated and compared with six classic instance segmentation models.(3)Edge detection and instance segmentation datasets for rock thin-section mineral particles are generated with meticulous annotations.Besides,basic information,data analysis results,and relevant evaluation metrics are presented in detail.Considering the different characteristics of the edge detection task and the instance segmentation task,a series of different pre-processing operations are adopted for the two datasets respectively,and the pre-processed results are shared with the community.