Research On Feature Representation And Similarity Learning For Visual Retrieval

Along with the upgrading of the multimedia technology,computer based visual reality technology has been widely applied.Specifically,new breakthroughs have been witnessed in various fields(e.g.industry,military,geoscience and the social life)by converting the objects in real world to computer-aided visual objects(e.g.3D model,3D&2D image).Moreover,the rapid development of Internet has enabled the transmission of visual objects resources regardless of their location.Under the condition of big data,the visual retrieval technology has become a hot topic,which would promote new technologies and products.Feature representation and similarity learning are two crucial issues of visual retrieval,which are also the focus of this thesis.The main contributions are summarized as follows:(1)First,on the basis of intrinsic space,global and local feature extraction methods are investigated.Spectral distance feature is extensively studied for global shape description,which has improved shape discriminative ability.Then,model heat feature is proposed and studied based on spectral distance and heat kernel,which has demonstrated superior performance than distance feature.For partial shape retrieval,the integrated heat kernel signature is further designed with much better performance than traditional methods.By theory and experiment,the studied global and local features have excellent properties,including intrinsic-invariant and noise resistence,which can help to improve the shape description ability.(2)Second,a novel multi-scale shape context approach is proposed to handle the scale offset problem which has never been considered before.Different from previous methods,this method is advantageous in: reducing the number of local features by keypoint detection while retaining the informative property of local featues,describing the spatial information by using multi-scale strategy,capturing shape distribution property gradually from the viewpoint of each keypoint and enlarging the feature space by discreting the the multi-scale feature for more complete dictionary learning.Experimental results show that the method can not only improve the shape matching accuracy,but also lift the performance of visual retrieval.(3)Third,a function transformation based similarity learning framework is proposed,and a new distance transformation based similarity learning method is developed.The main idea of this method is to define a new distance based on the similarity relationships of the dataset objects,which can discover the real similarity relationship between pairwise objects.Both the global and local models are considered for similarity learning.Experimental results show that the proposed approach can significantly lift the performance of the retrieval methods in the feature space,and it has achieved better results than most of the existing methods on standard benchmarks.(4)Fourth,concept ontology based multi-level deep learning approach is developed for large-scale visual recognition.The concept ontology is embeded into deep network to perform multi-task learning for 2D images which can be seen as the projection of 3D objects.The path selection property of the approach can provide a more detailed evaluation for classification.Besides the merits in time complexity,it can save up to 93.75% storage space compared with the baseline method by extracting group-specific feature.The experiments on large-scale datasets have demonstrated the superior retrieval performance of the proposed method.In all,this thesis follows the process of visual retrieval,where feature representation and similarity learning act as the key issue.Plenty of experimental results show that the proposed methods have achieved salient progress with respect to state-of-the-art and the other research related to visual retrieval can benefit from the proposed methods in this thesis.