Research On Visual Attributes Learning And Its Application In Image Retrieval

With the rapid development of modern communication and multimedia technology,the number of digital images on the network has increased rapidly,and how to find the human desired images quickly and effectively from the massive digital images,is a research topic with great theoretical value and practical significance.In order to narrow the semantic gaps between the low-level image features and the high-level semantics,how to automatically annotate the images based on the combination of low-level features and semantic visual keywords of the images,as well as how to perform efficient image retrieval using the learned image labels,has become a new research focus.In this thesis,we focus on the research of attributes learning and attribute-based image retrieval,the relevant research work has been carried out around three problems,and the innovation results are as follows:(1)For the problem of accurate representation of local features in relative attribute learning,in this thesis,we propose a deep relative attribute learning method based on attribute-correlated local regions.As many attributes are local-oriented,the local feature representations related to attributes can represent the change of attributes better than the global feature representations,hence,the method of discovering the spatial extent of attributes is improved in order to improve the accuracy of local location.Moreover,as the deep learning features are more powerful than the hand-crafted features,more accurate feature representations are obtained by concatenating the intermediate local features and the high-level global features extracted from the deep network.Experiments are carried out on the open image datasets,and the experimental results verify the effectiveness of the proposed method,which improved the relative attribute learning performance.(2)In order to solve the disorder problem of query attributes intensity in the binary attributes based image retrieval results,in this thesis,an image retrieval method based on grouped attribute strength is proposed.Through the original binary attribute-based method,the attribute strength of the images retrieved first,are often uneven,which cannot meet the user’s requirements.By combining the binary attributes and relative attributes to]earn the grouped attribute strength,the fine-grained elements are incorporated into the coarse-grained process,so that the images first retrieved have stronger attribute strength.Experiments are carried out on the open image datasets,and the retrieval results are more in line with the expected goals of the users,and have better retrieval performance than that of the original binary attributes based image retrieval methods.(3)In order to solve the accuracy problem of the semantic relation measurement,in this thesis,an image retrieval method using network search amount based extended attributes is proposed.The candidate extended attribute set is constructed by combining two external language libraries,and the user preference is measured by using the data information from the network,to eliminate the items that are not commonly used in the candidate set and obtain the final extended attribute set,according to the fact that the data from the network can directly reflect the usage and familiarity of users to the attribute words.When users conduct the image retrieval,the scope of the queries includes not only the pre-labeled attributes,but also the extended attributes.In addition,a consistency measurement method is also proposed to roughly verify the reliability of the learned extended attributes.The experimental results show that the proposed method,to some extent,improves the freedom degree of user query inputs and image retrieval performance.