A Research On Image Retrieval Based On Label Semantics

With the rapid development of Internet and computer technology,image data in the network has increased explosively.How to effectively store,manage and apply these image data is of great importance.In recent years,the emergence of various image based application scenarios poses great challenges to algorithms in the field of computer vision,including image retrieval.Although the existing image retrieval algorithms have good model performance on specific image data sets such as MNIST and CIFAR,the retrieval performance of the existing algorithms is greatly limited when the image scene is complex and diverse.The biggest difficulty in image retrieval is how to reduce the semantic interval between the underlying visual features of the image and the high-level semantics of users so as to meet the needs of users.In order to improve the performance of image retrieval algorithm,three aspects are studied in this dissertation:Firstly,in order to improve the efficiency of image retrieval,data sets are usually classified before image retrieval.Therefore,this dissertation proposes an image clas-sification algorithm based on the Latent Factor Model.On the basis of preprocessed formal image features and supervision information,the algorithm studies the semantic association between image features and categories by the LFM,transforming the image classification problem into the semantic learning problem of the latent factor learning.This model not only makes full use of the performance of LFM in latent semantic learn-ing,but also has higher learning efficiency than deep learning,makes the model have a wider application scene.Experimental results in two image classification scenarios verify the effectiveness of this algorithm.Secondly,in order to get richer semantic monitoring information,a deep multi la-bel image retrieval algorithm based on semantic graph model is proposed.Based on the idea of collaborative filtering and tag clustering,the algorithm learns the semantic asso-ciation between image multi label information,and generates the semantic monitoring feature for image.Then,we build a deep hashing model based on convolutional neural networks to learn the hash and eigenvectors of image to achieve fast and efficient image retrieval.The algorithm fully learns the rich user prior knowledge hidden in the image multi label data,getting more effective monitoring information for model.Moreover,the network model is built on the basis of existing work,which ensures the effectiveness of network in feature extraction and hash learning.Experimental results on benchmark image datasets show that the retrieval performance of this algorithm is better than that of existing models.Thirdly,considering the shortcomings of hashing representation in related works,this dissertation proposes an image retrieval algorithm based on semantic ranking.This algorithm constructs the optimization model by minimizing the distance between the im-age hash code similarity and the semantic similarity,and studies the semantic mapping and the potential association between the high dimension feature of image and the low hash code based on the optimal solution of the model.In this model,Hamming distance represented by Euclidean distance between hash codes is more reasonable and effective than the method of inner product calculation in existing work.In addition,the algorithm uses Natural Language Processing model Word2Vec to learn the semantic differences between different classes,so that this model can get richer semantic monitoring infor-mation.Finally,the effectiveness of the algorithm in different image retrieval scenarios is verified by comparative experiments.