Research On Image Comprehension And Retrieval Based On Deep Learning

Since twenty-first century,the level of Internet storage and computing capabilities have undergone a great leap forward,and the popularity of smart phones enjoys a substantially increase and users upload and share daily mass of image data.How users quickly find their own real information from the massive data?How to make the expression of image searching information more accurate and closer to the needs of users?How to integration large image data resources effectively and make it valuable has become a thorny problem in academic and industry.With the development of neural networks as well as deep learning,image comprehension is gradually understood from the image based on the underlying features(ie,using the visual features such as color,texture,shape and edge of the image to understand the image),to the high level of image semantic understanding(That is,the way in which human beings understand the image through the scene information expressed in the image).This thesis translates the underlying features that the computer can understand into the semantic features of human cognition,thus solving the problem of "semantic gap",and finally the resulting images retrieved from the computer are closer to humans based on understanding the retrieved image results.The contribution of our thesis is as bellows.Firstly,in the aspect of image comprehension,this thesis utilizes the advantage of deep learning to extract features by constructing neural network and extracts the visual features of images from images as well as the semantic features that traditional feature extraction methods can not extract to interpret images.Then,we propose a natural language processing method to solve the problem of text matching for the text description extracted for each image.At the same time,we introduce the Scene Graph with the method of natural language processing and propose a new Scene Graph matching method to solve image retrieval problems.In addition,we propose a new artificial annotation data set,and we test our methond on the data set to verify the superiority of our algorithm.