Multi Features Coding Based Image Retrieval

Image retrieval has been a hot research topic in recent years.Image retrieval systems based on various features and coding methods emerge one after another.However,because an image contains many types of data,it is difficult to interpret an image well with only a single feature.The multi-feature fusion image retrieval system emerges at the historic moment,which describes an image with a variety of features,and includes all kinds of information of the image,so as to achieve higher retrieval accuracy.Scale invariant feature transform(SIFT),Local Binary Pattern(LBP)and color feature are three commonly used features in image retrieval system,which have good performance in dealing with all kinds of images.Vector Quantization(VQ),Fisher Vector(FV)and Locality-Constrained Linear Coding(LLC)also have superior performance in encoding image features.In this paper,the three features and three encoding methods mentioned above are taken as the main research objects,and a variety of combination schemes are proposed for experimental verification in order to improve the image retrieval accuracy.The main research contents are as follows:1.Conduct experiments from the perspective of single feature and multiple coding,using one feature and three coding ways for combination.Firstly,a feature of the image is extracted,then k-means and GMM are used to generate the codebook.The features are encoded in the corresponding codebook in three ways to obtain the vector representation of the image.The cosine distance is obtained by corresponding the vector of the image to be retrieved with the vector of the database image,obtained three results.The final retrieval results are obtained by different combinations of the three results.Experimental results on the corel-1k dataset show that the combination of three coding methods with SIFT features has a better retrieval accuracy.2.Conduct experiments from the perspective of multi-feature single coding,using three features and one coding way for combination.Firstly,three features of the image are extracted,and then k-means or GMM are used to generate the codebook.The three features are encoded in the corresponding codebook in one encoding way to obtain the vector representation of the image.The cosine distance is obtained by corresponding the vector of the image to be retrieved with the vector of the database image,obtained three results.The final retrieval results are obtained by different combinations of the three results.The experimental results on the corel-1k data set show that the combination results of the three features have a better retrieval accuracy under FV coding.3.The experiment is conducted from the perspective of multi-feature and multi-coding,using three features and three coding methods for combination.Firstly,three features of the image are extracted,and k-means and GMM are used to generate the codebook.The vector representation of the image is obtained by encoding the three features in the corresponding codebook in three ways.The cosine distance is obtained by corresponding the vector of the image to be retrieved with the database image vector,obtained nine results.The final retrieval results are obtained by different combinations of the nine results.Experimental results on the corel-1k data set show that the combination results of the three features have a better retrieval accuracy under LLC coding and FV coding.