Image Retrieval Based On Perceptive Feature And Manifold Ranking

With the rapid development of visual images on the Internet and mobile terminal in recent years,visual image retrieval has been always the hot issue in computer vision field with potential commercial value.And the huge amount of images with complex visual contents bring great challenge for real-time image processing.Thus,image retrieval systems with high accuracy and efficiency is a major problem for urgent solution.Until now,content-based image retrieval systems aim to learn the similarity between the query and database using low-level visual attributes,in which image feature representation and similarity-based ranking paly important roles for retrieval performance.However,visual feature extraction and similarity-based ranking are independently learned in most of these systems,i.e.either extracting visual feature based on image content,or learning similarity measurement based on visual features.In this thesis,a novel image retrieval framework based on perception feature and manifold ranking is proposed by utilizing human visual manifold perception to consider the compatibility between image feature and ranking.Based on the analysis in human visual mechanism and Gestalt psychology for visual images,this thesis presents a novel global image representation named Perceptual Uniform Descriptor(PUD),which utilizes three principles in Gestalt psychology to detect the local neighborhood with similar visual attributes,and then describes these regions using two independent characteristics: contrast and spatial correlation.In the experiments,traditional distance-based measurement methods are applied to validate the effectiveness of this image representation.It is demonstrated that the retrieval framework combining L1 distance and PUD feature can obtain better retrieval performance than other measurements.Those results also illustrates that the compatibility between image feature and ranking plays an important role in image retrieval.Through projecting the PUD feature in 2D coordinate using classical manifold learning algorithms,the manifold structure among images can be visualized.Furthermore,this manifold structure is also the basic assumption in original manifold ranking algorithm.Based on the compatibility between PUD and manifold ranking,the retrieval framework in this thesis is presented.The discrimination of visual feature is the key to the improvement since manifold ranking is to re-rank the samples using adjacent graph based on the initial distance-based ranking results.And the manifold distribution of visual feature is able to further improve the retrieval accuracy.The extensive experiments in public image datasets validate the effectiveness of this retrieval framework.Also manifold ranking can improve the retrieval performance of other features which preserve the manifold structure.In addition,to solve the problem that the time complexity in manifold ranking is too high,a novel fast modified manifold ranking(MMR)is presented,which aims to update the small seeds similar to the query for similarity-based ranking.MMR utilizes the traditional manifold ranking to learn the similarity score between seeds selected from the database and the query,and then propagates the local score to all the samples in the database for re-choosing and updating new seeds.Therefore,MMR focuses on the similarity relation between seeds and the query,so small-scale manifold ranking greatly reduces the computational complexity.The experiments also illustrate that MMR can obtain higher retrieval accuracy with much lower computational complexity.