Research And Implementation Of Image Retrieval Technology Based On Multi-feature Fusion

Facing the rapid growth of image data in the Internet,how to retrieve interested images in the vast image collections accurately and quickly has become a hotspot of researching in the field of image retrieval.However,there may be some problems in the text based image retrieval,such as blurring and missing information around the image,irregular and multilingual image description,erroneously describing image and so on.Therefore,the demand of using image itself as retrieval input is increasing.Many content based image retrieval(CBIR)methods are proposed to extract visual content information of images,and improve the best matching accuracy of image features to be retrieved.Nevertheless,problems such as occlusion,overlap,difference of spatial layout and image resolution,illumination changes,semantic gap and exponential growth of multimedia content are making CBIR a challenging problem.In order to express the abundant visual information of images as fully as possible,this paper proposes an image retrieval framework based on the fusion of multi-source features,including the widely used deep convolutional neural network(DCNN)characteristics and traditional local or global features,which makes retrieval results more adaptive to users'demand and improves the accuracy searching.This article concretely expounds the following three contributions:1)Based on the improved "bag-of-features" model,the generation of coarse-to-fine visual dictionary and a new adaptive soft assignment technique are proposed in this paper.The framework integrate the "from global to local" concept in the construction of visual dictionary.The optimized product quantization technology compresses every local dimension of the visual dictionary obtained by unsupervised learning.In the stage of feature quantization,this paper controls the number of assigned visual words for each local feature and the weights of each visual word according to the proposed soft assignment mathematical model.Finally,the image representation of every single feature space is constructed.2)A multi-feature fusion method based on similarity weighting matrix is proposed.On the principle of feature similarity,the feature vectors of two similar images are very close in some dimensions and far away in certain dimensions.Therefore,the importance of similar dimensions should be strengthened and the contribution of different dimensions should be suppressed when calculating similarity.The multi-feature fusion method can dynamically weighted the features by using the similarity weighting matrix according to the difference of each dimension of features,which increases the weighting pertinence and improves the matching accuracy.3)A re-ranking algorithm based on visual correlation analysis of candidate images is proposed.Firstly,the initial retrieval results obtained by the existing image retrieval method are repeatedly searched.Secondly,the visual related direct graph among the candidate images is constructed according to the proposed method.Then,based on the proposed model,the visual correlation among candidate results is analyzed.Finally,the similarity weights among the candidate images are calculated,and the re-ranking is realized.4)An image retrieval system with the fusion of multiple features is designed and implemented.The users can choose the query image in the local databases to upload to the system.The system backstage will execute the image retrieval algorithm in this paper,and finally visualize the re-ranked retrieval results.