Feature Representation And Similarity Metric In Image Retrieval

Accompanied by the rapid development of internet technology,the ubiquitous visual data on the network presents an explosive growth trend with the popularity of various digital devices and embedded cameras.It makes the research on image search and retrieval technology more actively,and also brings new dawn to various emerging applications based on image search.Content-based image retrieval(CBIR)uses visual content as important clues for sorting images.It can effectively overcome the deficiency in visual information retrieval based on text,including inconsistent recognition between the text and visual content,weak ability for query intention resolution,and lack of user experience,etc.Over the past two decades,CBIR has attracted a lot of attention.Especially with the successful application of deep learning technology in the field of image process,the related research about CBIR is full of opportunities and challenges.Focus on two key techniques about feature representation and similarity measure in CBIR,we conduct some research from three aspects:how to extract and express the image features to imply the high-level semantic information as much as possible,how to effectively aggregate depth features to enhance the ability of expression with strong discrimination,how to define an exact similarity metric to reflect semantic affinity.This thesis includes the following aspects:(1)An image retrieval method based on multi-feature representation and diffusion pro-cess is proposed to improve the retrieval performance through image features expression and similarity metric.For traditional visual features representation,we design a model for feature representation that can effectively fuse multidimensional visual features such as color,texture,shape and bag of visual words(BoVW).From the angle of global and local,it can enhance the capacity of the expression of image characteristics through integrating lower and middle fea-ture,which effectively relieve the underlying differences between low-level visual feature and high-level semantic feature;At the same time,diffusion process is applied to our fused feature in image matching for optimizing distance metric.However,to relieve the limitation that the diffusion process might degenerate in measuring the distance of the top matches,a new search strategy is explored to further strengthen the retrieval performance.Some benchmark databases are used to evaluate the proposed model.The results of comprehensive experiments reveal that the enhanced representative capability of the proposed feature fusion schema and re-ranking based on diffusion processes can significantly improve the retrieval performance.Compared with related retrieval methods,the proposed method can achieve a higher retrieval accuracy.(2)An image retrieval method based on aggregated deep features is proposed.In view of the convolutional features based on convolutional neural networks(CNN),we put forward a kind of feature aggregation method based on regional significance weighting and channel sensitivity weighting,which can enhance the ability of description and discernability for deep features,so as to achieve the aim for improving retrieval performance.The proposed aggregation method can effectively take advantage of multi-scale,and assign different weight to different region according to the significance of contained visual content.At the same time,we design a sensitivity weight for different channel of deep feature based on the sparsity and intensity of response value in channel.Finally,the final image feature is formed through regional weighted aggregation.Although the proposed method only adopts the poo15 layer depth features of pre-trained VGG network model,a large number of experimental results on many benchmark image datasets including Holidays,UBK,Oxford5k,Oxford 105k,Paris6k and Paris 106k show that our method can achieve comparable retrieval results with state-of-the-art aggregation approaches of deep features without fine-tuned strategies and multiple image input.(3)An image retrieval method based on semantic feature representation and similarity met-ric is proposed.Different from the traditional practice,Axiomatic Fuzzy sets(AFS)is introduced to masterly embed the semantic into original image features for forming a new semantic feature space.In this new space different semantic description will be exploited for different image elements according to real data attribute distribution,which effectively reflect the semantic d-ifference between images.At the same time,in order to overcome the limitation brought by the traditional distance metrics,the similarity between not only any two images but also their respective neighbors are considered when we define the affinity between images.So,the orig-inal similarity between pairwise images is extended to the similarity between two image sets.Since the latent and stable data structure can be captured to make more accurate description and reflection of semantic affinity,it can further enhance the stability and robustness of similarity relationship of images.Extensive experiments on benchmark datasets verify the superiority of the proposed method.It has a good performance in the face retrieval task based on primitive pixel-level grayscale feature and the natural image retrieval task based on CNN features.