Image Retrieval Based On VLAD_CB

Human society with the progress of modern science and technology,has evolved into a highly developed society with Internet,and the Internet in people's lives mainly exists as multimedia which contains images and videos and many others.They bring us great convenience in the vast majority of leisure and entertainment.But what makes the image retrieval direction of the research personnel is how can we fast and accurate positioning target image to the user.In this big background,the image retrieval technology as a research hotspot in the field of computer vision,has made considerable progress.Its development is divided into basic text-based image retrieval and content-based image retrieval phase,the latter is more advanced methods,it has been widely used to now.At the same time,the content-based method also includes expression based on global features and expression based on local characteristics,which based on the expression of global feature is essentially based on local image characteristics of a sense of polymerization.Of this,Bo F(Bag of Features)is the most widely used because of its good effect,but because of it needs more clustering center,thus greatly increased the amount of calculation,the characteristics of feature points more images at the same time lose more information representation ability is abate,VLAD(Vector Of Locally Aggregated Descriptors)global express characteristics came into being,the characteristics of the statistics of the image in a local features and their clustering center of Euclidean distance information,compared with Bo F reduces the calculation time also obtained the higheraccuracy,but to some extent the algorithm still exists the disadvantage that lack of descriptor number information,affect the image retrieval ability.Aiming at the problem,this article start on how to enhance the capacity of VLAD characteristics description,proposed a new global express characteristics VLAD_CB(Vector of Locally Aggregated Descriptors Combine Bag of Features),this algorithm combines VLAD and Bo F advantages,in view of the high efficiency and precision of VLAD and the comprehensiveness of descriptor of Bo F,the synthesis optimization is done to a certain extent.Bo F statistics to image local characteristic clustering after description of the amount of information,VLAD_CB not only consider the descriptor classification and clustering center after the distance of the problem,at the same time will also joined the quantity information,the more abundant information of global features.Basic idea is as follows:1.Using SIFT(Scale Invariant Feature Transform)partial feature of the feature extraction and image library,will receive the image library for K-means clustering.2.To obtain the clustering center number K,forms K word table,statistics the number of occurrences of each word in the image,and image expression into a K(K × 1)dimensional numerical vector.3.After VLAD calculated K × 128 dimensional feature vector,the two combine to form K × 129 dimensional vector.In this paper,through calculating the m AP(Mean Average Precision)value method,the greater the value represents the calculation method has better efficiency.VLAD_CB method not only retains the advantages of VLAD,at the same time increased the Bo F numerical information,better solved the high cost of computing and weak characterization.By the representative of the INRIA Holidays database and Oxford5 K database after the experiment,the experimental results show that VLAD_CB algorithm is compared with the traditional characteristics of Bo F and VLAD showed more accurate data,with little more storage.After theoretical analysis and experimental results of this paper,show that if no extreme conditions appear,VLAD_CB will bring the user experience better,to verify the effectiveness of the algorithm of image feature extraction.