Research On Pore Detection Of Sandstone CT Image Based On AT-Mask RCNN

The pore detection of sandstone CT images is an important task in geology,and it has important guiding significance for studying the internal structure of sandstone and the geological information of sandstone reservoirs.The traditional method is mainly to manually identify the pores of sandstone,which is time-consuming,labor-intensive,subjective,and the identification standards are not uniform.The development of automated detection methods for pores in sandstone CT images helps to save manpower and improve identification efficiency.The core technology of automated detection methods is target detection,so the pore detection in sandstone CT images based on deep learning is of great significance.In the process of target detection,there are often situations in which the targets to be detected are adjacent,close,or occluded,and there is a problem of missed or repeated detection of the target,and this problem also exists in the task of pore detection.Aiming at the above problems,this paper improves the Mask RCNN algorithm based on candidate regions,and applies the improved algorithm to pore detection in sandstone CT images.The main research contents are as follows:1?Aiming at the target missed detection and repeated detection problems caused by the fixed threshold in the non-maximum suppression algorithm of Mask RCNN,an improved Mask RCNN algorithm based on AT-NMS is proposed.First,add a deformable convolution module to the Res Net50 network to enhance the ability to extract multi-layer convolution features of the target.Secondly,use the second-order difference to calculate a reasonable threshold,and extract the deep information of the target candidate region in the RPN stage,and then process the pair by ROI Align The positioning of the target is more accurate,and finally the target instance segmentation,target classification and target frame regression are realized through three branches.2?Perform image preprocessing on the collected sandstone CT images,and self-generate a pore detection data set.First,perform image preprocessing operations such as image rotation,image denoising,and random interception on the collected sandstone CT images to expand the data set.Then,under the guidance of relevant experts,the pores of the sandstone CT image are marked.Finally,an annotation data set of sandstone CT images is obtained,which lays the foundation for the pore detection task.3?Aiming at the feature visualization and performance issues of the pore detection model based on AT-Mask RCNN.The feature visualization operation will be performed on the intermediate activation layer and the module convolutional layer to show the model learning process.On the sandstone CT image test data set,a comparative experiment and statistical analysis of the Mask RCNN algorithm and the AT-Mask RCNN algorithm are carried out,and the overall performance of the model is evaluated using two evaluation indicators,namely,pore segmentation evaluation and pore counting evaluation.