| Since deep learning has been used in the field of object detection and recognition in images,breakthrough progress has been made.It has become the mainstream method in the field of image recognition.More and more object detection frameworks based on deep learning have been proposed,and they have been industrially applied in various fields.The technology of object detection and recognition is introduced into the scanning electron microscope image of cement,and the SEM image of cement is analyzed by means of computer science.This thesis is based on the research topic of the application research of convolutional neural network in cement SEM image object detection.The Faster R-CNN object detection framework is selected to detect the cement SEM image.Aiming at the characteristics of hydration products and pore sizes in cement SEM images that have large differences and overlap each other,the improvement of Faster R-CNN improves the detection accuracy,and provides ideas and methods for the subsequent study of cement microstructure.The main work done in this thesis is as follows:Aiming at the problem that the cement SEM image data set is too small in this thesis,the characteristics of each convolutional neural network are analyzed,and a Faster R-CNN based on DenseNet feature extraction network is proposed.Compared with other convolutional neural networks,DenseNet has better anti-overfitting ability and generalization performance.It can effectively extract features from cement SEM images and reduce overfitting.Experiments show that compared with Faster R-CNN object detection model based on commonly used convolution neural networks,mAP has been improved to a certain extent.Secondly,in response to the problem of large size differences between objects in the object detection task in this thesis,an adaptive anchor size design method is proposed.The clustering method is used to cluster the normalized length and width of the object in the object sample of the cement SEM image.Clustering is based on the intersection over union(IOU)of the boundary box of the object as the distance of clustering analysis,at the same time,the area of the object's bounding box is statistically analyzed to obtain the size and ratio of the new anchor,which is used as the parameter in the RPN,which improves the accuracy of object detection and reduces the missed detection caused by the default anchor size.In view of the fact that some objects in the cement SEM image are too small and some objects overlap,which leads to the problem of missed detection,this thesis introduces ROI Align in Faster R-CNN.ROI Align can reduce the impact of quantization on small objects during feature extraction.In general,this thesis improves the Faster R-CNN with higher accuracy in the current object detection framework,and then detects the hydration products and pores in the cement SEM image.The experimental results show that the improved Faster R-CNN object detection framework can detect objects in the cement SEM image,and the mAP is increased by 8.93% compared with the unimproved Faster R-CNN based on VggNet as feature extraction network. |
