Research On Large-scale Image Retrieval Algorithm Based On Features

With the rapid development of information technology,various Internet products are rapidly gaining popularity in people's lives.The development of information has brought about an increase in data,especially the image databases.In the face of large-scale image database,in traditional text-based image retrieval,there are problems of complex manual annotation,inaccurate description of text,low accuracy of retrieval and time efficiency.In order to satisfy people's request,researchers have proposed the content-based image retrieval,that is,to retrieve the image according to the information of the target itself.The traditional large-scale image retrieval algorithm has low accuracy and slow speed.For the problem,this paper innovates the algorithm based on the content of the image.The innovations are as follows:(1)To solve the problem of long construction time and low accuracy of the visual vocabulary tree in traditional large-scale image retrieval,this paper proposes the visual vocabulary tree based on CICDMDTC(Clustering by Inter-Class Dispersion,Maximum Distance and the Tightness of the Class).Firstly,in view of the cluster definition,the clustering algorithm CICDMDTC on the basis of inter-class dispersion,maximum distance and the tightness of the class,is proposed to solve the instability and low exactness caused by traditional K-means' sensitivity to the initial clustering center,the selection of K and the noise points.It is proved that the method has higher accuracy and stability.Then a vocabulary tree based on CICDMDTC is proposed.This method constructs a vocabulary tree by clustering hierarchically on the extracted features.In the end,SURF is used for feature extraction to implement large-scale image retrieval algorithm SURF_CICDMDTC.Compared with the experiments based on the traditional visual vocabulary tree algorithm and other improved methods,the results show that the time of constructing vocabulary tree is shorter and the retrieval algorithm based on the improved method of this paper is improved in accuracy and time efficiency.(2)In order to solve the problem that the features extracted in traditional algorithms fail to fully represent the image content,which results in low retrieval accuracy,this paper proposes the feature extraction algorithm named SURFH_PCA.The method firstly uses the integral image and the Hessian matrix to determine the location and scale information of the feature points,and then extracts the 64-dimensional Haar features and the48-dimensional HSV color histogram features in their neighborhoods,and finally normalizes them.Using the principal component analysis respectively to reduce the dimension of the vector,we obtain the 56-dimensional SURFH_PCA feature vector.(3)For the problem that the traditional algorithm can not balance retrieval accuracy and time efficiency,this paper proposes the SURFH_PCA_CICDMDTC large-scale image retrieval algorithm.The algorithm uses SURFH_PCA to extraction image features and uses the CICDMDTC clustering algorithm to manage the retrieval process.Experimental results show that the method has higher accuracy and time efficiency.It is illustrated that the improvement is effective,which makes it possible for people to retrieve targets from large-scale image sets fast and accurately.