Research on Segmentation of Sandstone Thin Section Images Based on Semantic Identification

Author:Jiang Feng

Supervisor:Gu Qing


Degree Year:2018





The sandstone thin section microscopic images,or sandstone images for short,are made from sandstone samples which are ground into thin sections with the thickness of 0.03mm.The thin sections are put on the stage of the polarizing microscope where the images are captured by the high-definition camera.In the research area of geology,the identification of sandstone thin sections is to study and analyze the compositions and percentages of minerals within the sandstone images,and further perform sand-stone classification and naming.The identification of sandstone thin sections plays an important role in oil and gas reservoir exploration and evaluation,environmental protection and water conservancy,etc.The segmentation of sandstone images is to partition the mineral grains into separate regions within the thin sections,which is the first step and a primary work of the identification of thin sections,and has been of great significance.The sandstone images have their own characteristics,i.e.,the image contains a large number of mineral grains,the contrasts between adjacent grains are low,and the micro-structures inside the mineral grains are complex,which make the segmentation of sandstone images more difficult than the ordinary images,and hence it’s a chal-lenging work.In this dissertation,based on the principle of semantic segmentation,we extract the semantic information contained in the images and research the segmentation techniques of sandstone images.Our work and contributions include the following:(1)A framework for segmenting multi-angle cross-polarized sandstone im-ages is proposed.In this dissertation,we propose a three-stage framework for segmenting multi-angle cross-polarized sandstone images,which is composed of a pre-segmenting stage,a feature extracting stage and a region merging stage.In the pre-segmenting stage,considering the relationships of multi-angle images made from the same sandstone thin section and its geological meanings,we propose the MSLIC(Multi-angle SLIC)algorithm,which partitions the input image into homogenous regions composed of neighboring pixels sharing similar properties,i.e.,the set of superpixels.In the feature extracting stage,we adopt the advises from the geologists,and combine the brightness features,the texture features and the boundary features to represent the superpixel-s.Meanwhile,we define the feature similarities between superpixels according to the properties of different features.In the region merging stage,the RAG(Region Adja-cency Graph)is constructed according to the superpixels and the similarities among them,which are obtained in the pre-segmenting stage and the feature extracting stage,respectively.And then,the multi-angle region merging method MRM is performed on the vertices of RAG to merge the over-segmented mineral superpixels.Furthermore,we propose a globally profit function for evaluating segmentation results which are used to automatically stop the merging processes,in order to obtain the optimal merg-ing results.The experiment results show that the proposed method is more suitable for segmentation of sandstone images comparing with other image segmentation method-s,for the reason that it can segment hundreds of objects contained in the image into independent and complete mineral grains.(2)A handcrafted features extraction based method for segmenting sand-stone images is proposed.The low-level features of an image contain the local features of single pixel or its neighboring pixels,which are difficult to describe the semantic information of the sandstone images.Therefore,the image segmentation method based on low-level fea-tures is not effective for sandstone images with complex compositions.In the feature extracting stage,the handcrafted features are designed to represent the image semantics according to the characteristics of sandstone mineral grain images.In this dissertation,the histogram-based color features,the multi-frequency and multi-orientation texture features and the multi-scale boundary features are combined to characterize the super-pixels.Hence,the combined features can identify the category of the mineral grains,and distinguish different grains of the same category as well.Furthermore,in the re-gion merging stage,we propose an efficient and accurate two-step superpixel merging method,called CoFM(Coarse to Fine Merging),based on three rules of Gestalt Cog-nition Theory,i.e.,the law of proximity,the law of color constancy and the law of similarity.The proposed method merges the "unambiguous" superpixels by checking the boundary features at first,and then merges the "ambiguous" superpixels by recog-nizing and clustering their color features and texture features.Finally,we prove that the convergence of solution sequences generated by our algorithm can be guaranteed by Zangwill theorem.The experiment results show that by extracting the handcrafted features,the image semantics can be identified and the categories of superpixels can be recognized,thus the mineral grains of different categories in the sandstone images can be correctly segmented,which outperforms the performance of the-art-of-state image segmentation algorithms.(3)An automatic feature extraction based method for segmenting sandstone images is proposed.Handcrafted features dependent on the domain knowledge of designer and cannot fully characterize the essences of sandstone mineral grain images.Therefore,in this dissertation,we propose an automatic feature extraction method to acquire the features of mineral superpixels.By analyzing the optical properties of mineral grain images and studying the typical convolutional neural network models,we design the convo-lutional neural network named RockNet,which is used to automatically extract the features to characterize the superpixel semantics.Training of RockNet needs a large number of samples,however,the number of labeled sandstone grain samples is limit-ed,hence the transfer learning and the multi-angle images augmentation methods are used to learn the network parameters.In addition,we propose a method to speed up feature extracting processes by using entire image convolution and featuremap map-ping.In the region merging stage,an improved region merging method called FCoG(Fuzzy Clustering of Rock Grains)is proposed by mining the relationships between adjacent mineral superpixels,which guarantee the image segmentation result to be the globally optimal solution.The experiment results show that the features extracted by RockNet are better than handcrafted features in describing the surface properties of mineral grains and the micro-structures,e.g.twinning,cleavage and alteration.Fur-thermore,the FCoG method can accurately cluster and merge the mineral fragments which are hard to recognize,by exploiting their features and the relationships among them.Finally,the ideal image segmentation is obtained.