A Retrieval Method Of Image Time-series Similarity Based On DTW

The arrival of mobile internet era is speeding up the sharing and flow of information.However,too much information has caused information disasters and the multimedia resources are suffering as well.Nowadays,multimedia are developing as one of the major medium of internet information dissemination,among which images are the most simple and intuitive carrier of multimedia resources and this has become a hot field in recent years.However,the traditional content-based image retrieval technology are facing the two primary problems;One is the difficulty of abstracting feature vectors,the other is the lack of efficiency with sequential retrieval model.The high-dimensional indexing techniques are suffering from the problems,such as time consuming during indexing and the high frequency of index updating.The technique with problems above mentioned can not meet the demand of current quick search of images.Therefore,a new retrieval method is in urgent need to solve the problems.In this paper,we combine the concept of image time series to explore the field of image retrieval,aiming at solving the problems above mentioned.The main work includes:1.To solve the problem of losing high dimension vector information,an improved method of text-expression of high dimension vector was proposed.Commonly,before applying the full-text retrieval technology to the process of high dimension vector retrieval,high dimension vector was converted into text information and the simplest way is to convert every number in high dimension vector into string.However,the high dimension vector may lose some information during the converting.This paper improved the method with appending a dimension mark to each string,which represents the high dimension vector information and solves the problem of losing information.2.The main work of this paper is proposing a retrieval method of image time-series similarity based on DTW,which is introduced from voice recognition field.This method combines the time-series similarity and the advantage of dynamic time warping algorithm.First,the images are traversed and the pixel sequence is abstracted with the feature of time series.Then,a set of RGB color data is appended to the pixel sequence and the number sequence is generated in this way.Finally,the similarity distance is calculated with DTW algorithm.Experiment results show that the proposed method is able to quickly retrieval similar images.Compared with the traditional content-based image retrieval technology,additionally,he proposed method does not need to abstract the feature vector of image and index every target.What is more,the method is able to retrieval similar images with low time consuming and high accuracy.