Research And Application Of Image Retrieval Based On Quantitative Learning

With the rapid development of the Internet,the new picture data is measured in hundreds of millions every day.However,among so many images,it is crucial to find the ones that are valuable to you.In the field of image retrieval,image retrieval based on hash learning and image retrieval algorithm based on the quantitative characteristics of image can be transformed into low dimensional binary code,but the hash directly after compression can lead to information loss is too large,according to the binary code retrieval accuracy is not high,the problem such as the distribution of the image characteristics depend on the big,but based on the quantitative study of image retrieval can overcome these problems,this is undoubtedly in direction a huge improvement.Image retrieval algorithm based on quantization is more and more researched and applied due to its higher information preservation and accuracy than hash,but it also has the advantages that hash does not have to restore the original image through binary code.Early image retrieval is generally supervised,that is,users are required to annotate the image data and a paragraph of additional supplementary text is required to describe it.Although the result is very good,it is very time consuming and material cost.Considering the increasing speed of image data,this is obviously unacceptable.Industry and academia prefer unsupervised image retrieval algorithms.In order to solve the above several deficiency,in this thesis we put forward the following three:(1)The first quantitative study image retrieval algorithm on the compression of a searchable database of a rough quantitative screening algorithm based on k-means algorithm,used in the pretreatment stage to cluster the database,we according to the query in retrieval phase selected cluster to form a new smaller retrieval scope,the experimental results show that our algorithm can greatly narrowed the retrieval calculation so as to improve the speed of retrieval.(2)Then this thesis presents a new unsupervised MCQ quantitative learning algorithm of high recall rate.The experimental results show that: 1)the introduction of the candidates to speed up the encoding process,the speed of the improved the coding accuracy;2)minibatch gradient descent to join to make iterative loss lower limit is lower and more stable;3)unconstrained code and binary code generation after decoding makes features closer to the original;4)increased at the same time,we use the dropout to prevent code book may lead to the fitting,the experimental results show that our method to more than all the quantitative unsupervised image retrieval.(3)Finally,we put forward a kind of based on hierarchy quantitative MCQ algorithm,the algorithm based on the code of the hierarchical structure of code this update and the coding phase,makes the code this can still be in the absence of orthogonal constraint of low computational time complexity of updating and lower consumption of coding.