Image Retrieval Based On Graph Model

With the rapid development of Internet technology,multimedia data shown an exponential growth trend.How to mine useful key information from huge database is an urgent problem to be solved.Image retrieval task focuses on image data,tar-geted at finding image sets from image database which satisfy the intension of users well.Moreover,content based image retrieval(CBIR)allows users to directly launch query by virtue of image,therefore,the visual characteristics and semantic attibute have become quite hot research topic.In the pipeline of CBIR,image representing and sim-ilarity measurement are of great importance to the performance of retrieval in terms of precision.Given image representations(features),similarity scores between images can be efficiently determined under euclidean distance metric,which is widely adopted in image retrieval.However,a lot of work has demonstrated that the capacity of euclidean distance metric is limited by the discriminative power of image representation,due to lack of employment of image distribution and neighboring relationship in feature space.In this thesis,we concentrate on CBIR and carry out research on graph model,and we propose two methods:cascaded feature augmentation based on diffusion and graph neural networks for feature re-representing,to integrate information from nearest neighbors of images to their representations,thus producing new features with greater discriminative power and further significantly improving the capacity of euclidean dis-tance metric.First,we propose cascaded feature augmentation based on diffusion(CFA).Diffusion-based methods for image similarity measurement is widely used in re-ranking stage of online queries,which have limited time efficiency and spatial efficiency.Based on the above consideration,we propose to migrate the diffusion process from online stage to offline stage,and utilize it to mine the manifold structure and potential correla-tionship of database images.Then,we linearly aggregate the features of adjacent nodes based on the mined neighboring relationship to form a new representation for each im-age in database.Further,in order to make full use of the hierarchical characteristics of neighboring relationship,we conduct feature augmentation in multiple scales.In this way,the manifold structure is embedded into the image features,so that potential mutu-ally related images are getting closer in feature space.In other words,implicit clustering effect is achieved.To scale the diffusion method to large-scale image database,we pro-pose a cascaded cluster diffusion technique for feature augmentation which largely re-duces quantization error.Finally,we extend our cascaded feature augmentation scheme to cases with multiple features which are complementary.Second,we propose graph neural networks for feature re-representing(GFR).Graph neural networks(GNN)are highly effective in tasks like semi-supervised clas-sification,however,GNN have not yet received enough attention in image retrieval task.In this thesis,we utilize GNN to learn an aggregation function,which is ap-plied to fuse the features of neighboring nodes to re-represent each image with sup-port of weighting mechanism.By combining residual connection fashion,graph model with different depths can be quite easily constructed through stacking modules,which are specialized to make use of neighboring relationship in different degrees.We only require a small quantity of training samples to gain excellent graph neural networks in re-representing images via jointly optimizing classification and ranking objectives.Based on the new image representations,similarity scores of images within same cat-egory are further boosted,while the distances between images of different categories are expanded.Moreover,we put forward an approach termed self neighbor to generate new features for query images,which is very convenient and efficient.The proposed method can flexibly adapt to different sizes and different forms of graphs.In addition,our method can be seamlessly combined with other re-ranking methods to further im-prove retrieval precision.Experiments on several public retrieval datasets demonstrate that methods pro-posed in this thesis are featured by high precision,low retrieval complexity and good scalability.