Research On Image Retrieval Based On Multi-feature

As one of the most important parts of the Internet data,the image data has been accumulated at an alarming rate with the rapid development of Internet information era and the wide range of smart phones.Compared with the text data,the advantage of image data is that it provides richer and more intuitive content information.So how to achieve effective organization and management of the images in a large image database in order that people can fast retrieve and access to the required images has become an increasingly important and challenging research problem in the information age.The earliest image retrieval systems are mostly based on text retrieval,and their retrieval performances depend on manual annotation,but the manual annotations of image are subjective and cannot express all the information contained in the image.Compared to the traditional text-based image retrieval system,content-based image retrieval system avoids manual annotation,but it cannot fully meet the needs of practical application,either the robustness is not strong,or the retrieval efficiency is too low.In nearly a decade,with the efforts of the universities and scientific research institutions at home and abroad,the excellent local features(e.g.,SIFT)and the bag-of-visual-words model were proposed,and their application greatly promotes the content-based image retrieval to a higher level of development.In this dissertation,a new image retrieval method based on multi-feature is proposed.The proposed image retrieval method focuses on the two-dimensional inverted index embedding semantic attributes,which can make the candidate images,not only sharing a large portion of similar local features but also being consistent with the semantic attributes,rank higher in the retrieved set.This dissertation introduces the theory of content-based image retrieval and analysis the traditional inverted index and the extraction of semantic attributes,and then proposes the new image retrieval approach.The main stages of the proposed approach are divided into two parts which are building the two-dimensional inverted index and updating the two-dimensional inverted index.In the stage of building the two-dimensional inverted index,the image features,i.e.,the SIFT and Color Names(CN),are extracted based on scale-invariant keypoints which are detected with the DoG(Difference of Gaussian)detector.Then,we use the two codebooks,which are trained using independent SIFT and CN features,and quantize the image features SIFT and CN,respectively.After that,each of the keypoints of the images can be represented by a visual word pair and then build the two-dimensional inverted index for the image database.Two-dimensional inverted index is equivalent to the fusion of two kinds of visual information on the index level,which can greatly enhance the precision of visual match.In the stage of updating the two-dimensional inverted index,we firstly convert the semantic attributes to probabilistic representation and employ the Total Variance Distance(TVD)to measure the semantic distance between the probabilistic representations.Then,the updating algorithm will search for the top nearest neighbors of database images using the semantic attributes and insert them to the two-dimensional inverted index.The updating method can embed the semantic attributes into the two-dimensional inverted index,so that the performance of image retrieval can be improved to some extent.In experiment part,the proposed image retrieval method based on multi-feature is validated in Ukbench dataset and Holidays dataset,respectively.The experimental results show that the proposed image retrieval approach based on multi-feature can obtain better retrieval performance,and has a certain application value in the field of near-duplicate image retrieval.