Research On The Appliction Of Deep Learning In Image Retrieval Of Rock Slices

As an important part of rock composition,rock pores are the basis for studying reservoir migration,storage mechanism and its control.However,different types of pores have a large effect on oil and gas permeability.Therefore,this paper uses pore images of the tight reservoir sandstone in the Ordos Basin as experimental data to study three image retrieval methods.Firstly,color features,shape features and texture features are used as starting points.Content-based image retrieval techniques are applied to rock pore images retrieval.The rock pore images retrieval based on color histogram,SIFT shape feature and Gabor wavelet texture feature is realized by experiments.The results show that the method has low recognition rate of rock pore images.As a result,the retrieval efficiency is not high,and the average is only50%.Secondly,the parameters of the pre-training network VGG16 are fine-tuned and used to extract and retrieve the features of the rock pore images.Compared with the former method,the rock pore images retrieval efficiency based on the pre-training network VGG16 has improved,reaching an average of 70%.But it still cannot meet the actual needs of search.Finally,the paper selects the AlexNet network model,and adjusts the number of convolutional layers,convolution kernel size and learning rate of the network according to the sample dataset.So the network structure that is most suitable for rock pore images recognition is obtained.The rock pore images retrieval experiment was completed by the network structure.The experimental results show that the adjusted AlexNet model is more efficient in rock pore images retrieval,with an average of more than 93%.