Research And Application Of Rock Image Retrieval Based On Multi-granularity Networks

In recent years,with the rapid development of convolutional neural network and its excellent performance in large-scale image processing,researchers in many fields have turned their attention to deep learning.The combination of geological science and computer technology is a new research topic and research direction.For practitioners in the geological industry,how to classify and manage a large number of rock sample images in the information age and how to effectively retrieve them are worthy of research and application.Although deep learning has performed well in the field of image retrieval,there are still many problems in its application in the field of geology,such as high computational complexity,difficult to apply in real time,network training requires a large number of data to complete,and it is difficult to apply in this field.In order to solve these problems,this thesis preprocessed and enhanced the existing data,using the idea of multi-granularity network,using the information extracted from multi-granularity feature map to improve the performance of recognition and retrieval,which can effectively improve the problem of insufficient data.The research content of this thesis mainly includes the following points:First of all,in view of the problem that the amount of rock sample image data is insufficient at present,this thesis uses limited data to enrich the number of data and scenes through data preprocessing and data enhancement,divides the training set,test set and verification set,and manually annotates the pictures in the data set.Secondly,because the fine-grained features of rock samples are of great significance to the recognition of rock sample image,this thesis uses multi-granularity network to extract the features of different granularity,and optimizes the parameters according to the characteristics of the application field.According to the results of multiple sets of comparative experiments,it is shown that the precision of rock sample image retrieval based on multi-granularity network is significantly improved,which shows that fine-grained feature is effective and necessary for rock sample image recognition,and multi-granularity feature extraction is helpful to improve the recognition ability of the network for rock.Finally,in the above methods,the data source needs to be manually labeled or the target in the image is extracted by the detection algorithm to process the image.In this thesis,a method of rock sample image retrieval based on aligned multi-granularity neuralnetwork is proposed.By combining the spatial transfer module to learn the information of the target itself,the input data can adjust the size of the target in the image by self-learning Small and position,as well as angle,make the model improve the recognition effect of various targets,especially small targets,and also improve the robustness of the model.The experimental results show that the proposed method is effective.