Deep Shale SEM Image Segmentation Based On Convolutional Neural Network

With the successful exploration and development of shale gas in the United States,there has been an upsurge in shale gas research around the world.Although China is rich in geological resources,the geological conditions are complex,and the tectonic evolution,sedimentary environment and thermal evolution process are different in different regions,which make the formation process and enrichment degree of shale gas vary greatly.Therefore,exploration technology is still the key factor in the development process of shale gas.By using image segmentation technology to segment SEM images of deep shale,it can provide auxiliary information for exploration personnel,which can undoubtedly increase exploration efficiency.In order to improve the accuracy of segmentation,this paper conducts an in-depth study on the segmentation of deep shale SEM images.The organic matter such as kerogen can be segmenting and extracted by image segmentation technology,and the exploration personnel can judge its growth state according to its morphology information.At the same time,explorers can infer the degree of shale mineral development from the information about the proportions of different fractions,and thus determine the degree of shale gas enrichment.However,through experiments,it is found that if the general image segmentation model in the field of computer vision is directly applied to the deep shale SEM image for segmentation,the final result is not good.Because the difference of light and shade in the gray image will cause the contrast difference between the targets,it will affect the effect of feature extraction.In this paper,the gray-scale image enhancement algorithm is used to preprocess the original image.While adjusting the overall brightness difference of the image,it also enhances the contrast between different segmentation targets.We first used the classic instance segmentation model Mask R-CNN to segment the image,but the segmentation accuracy was insufficient.Therefore,we use the idea of DFN network to improve the model and enhance the ability of extracting edge features.The improved branch merges feature maps of different levels as input,and uses up-sampling to restore the image,which effectively improves the model's detailed feature restoration ability.After that,the segmentation accuracy can be improved.By using multi-gray threshold segmentation,different substances can be segmented from deep shale SEM images.However,the gray values of clay and matrix mineral particles are similar,so other methods are needed.Since clay tends to grow together with organic matter,these mixed substances are interlaced in the image and have strong texture features.The U-net model is widely used in the medical field because of its good texture feature extraction capabilities.Therefore,based on the U-net model,this paper improves the loss function in the training process and uses the weight based on spatial feature information to enhance the comparison of boundary features,so as to improve the ability of U-net to distinguish clay mixture from matrix mineral particles and achieve the acquisition of different material proportion information in the SEM image of deep shale.In this paper,experiments on real deep shale SEM image data sets prove the effectiveness of the improved method.By comparing with the unimproved model,it is proved that the improved method can indeed improve the segmentation effect,which has certain practical significance and application value.