Research On Image Segmentation And Recognition Of Sandstone Thin Section

In the field of geology,the identification of sandstone thin section has important guiding significance for oil and gas exploration.At present,manual observation is mainly adopted in the analysis of sandstone thin section,which consumes a lot of costs,and the identification quality is inconsistent due to different analysts' experience.The development of an automatic identification device for sandstone thin section helps to improve the identification efficiency.The core technology of the automatic identification device is to identify and analyze images,so it is of great significance to research the image recognition technology of sandstone thin section.In this paper,the sandstone thin section are processed by two steps:segmenting first and recognizing second.First,the sandstone thin section image is segmented based on the superpixel algorithm,and then the convolutional neural network model is trained to identify the component categories in the sub-image one by one.Therefore,this paper focuses on the two parts of superpixel segmentation algorithm and image recognition model.The main research contents and results are summarized as follows:(1)For the segmentation of sandstone thin section images,this paper proposes an adaptive superpixel SLIC algorithm(AS-SLIC),which can dynamically generate superpixels based on the regional color histogram.While improving the superpixel algorithm characteristic distance metric function,it is capable of binding a plurality of images generated superpixel.Experiments on sandstone thin section images prove that the boundary recall of the proposed algorithm is improved compared with the SLIC algorithm,and the multi-image segmentation algorithm further improves the precision compared with one single image,which alleviates the under-segmentation problem of the traditional algorithm.(2)For the problem of over-segmentation,this paper proposes a superpixel merging algorithm based on the similarity of regional features.This algorithm extracts regional features and merges over-segmented regions according to the feature distance between regions.Experimental results prove that the proposed superpixel merging algorithm can alleviate the problem of over-segmentation,can greatly improve the precision of boundary segmentation,and make the segmentation result closer to the result of manual annotation.(3)For sandstone thin section image component recognition problem,this paper designs a lightweight convolutional neural network model.In this paper,the Res2Dw model is proposed,which combines the ideas of depthwise separable convolution,dense connection,residual learning together,and reduces the amount of parameters while giving the model a deep network structure and multi-scale receptive fields.At the same time,strategies such as image enhancement and class balance loss function are used to alleviate the impact of data distribution imbalance.The experimental results show that Res2Dw is a lightweight model which has higher recognition accuracy than models of the same magnitude.(4)As for model recognition accuracy,this paper verifies through a large number of experiments.In the research process,two image datasets are made and explored to improve the accuracy of model recognition.The final trained model can identify 105 kinds of sandstone components with an accuracy rate up to 89.8%.