The Application Of Feature Extraction On Image Searching And Near-duplicate Image Elimination

In recent years,with the vigorous development of information technology and Internet application,the multimedia data are growing at incredible high speed that we could not imagine.As a kind of indispensable information carrier,digital images have a very important role.Because digital images have some redundancy,it is forming an important research field that how to use the image processing techniques such as image feature extraction to manage the growing image database effectively.In this paper,we have studied the research area of the image searching and near-duplicate image elimination according to image feature extraction technique.An image searching algorithm based on bag-of-words(BOW)and a hierarchical near-duplicate image elimination algorithm have been proposed.For the image searching algorithm based on bag-of-words,in general,we extract features from an image and searches it by a combination manner of rough matching and fine matching to balance the speed and precision.The main contributions of this algorithm are as follows: Firstly,the traditional BOW is improved by adding some information of spatial relationship.In this way,an image is divided into three circular parts: the inner,middle and outer parts.So,the searching precision is improved while the searching speed is still fast.Secondly,a feature point extraction method on concentric circles is proposed,which is resistant to geometric attacks such as image translation,rotation,scaling,and adding noise.This method can make full use of the idea of statistics of BOW to extracted feature points evenly and fleetly.Thirdly the Gabor local line-based feature(GALIF)descriptor is improved by changing the filter objects from the lines to small regions.Through this method,we solved the problem of low efficiency and rotation-sensitivity.What's more,a new image signature matching strategy is proposed based on Euclidean distance and a scale invariant feature to improve the matching accuracy.Finally,the most similar image will be found using the Oriented FAST and Rotated BRIEF(ORB)method as the results of the fine matching step.Experimental results on a database with thousands of images show that our algorithm can retrieve mobile images quickly and precisely.What's more,we have studied the near-duplicate image detection and elimination algorithm.A hierarchical near-duplicate image detection and elimination technology is proposed which consists of “harsh grouping” and “deep grouping” to divide an image database into groups.The main contributions of this algorithm are as follows: Firstly,a global descriptor is proposed based on the density of feature points which describe density information of the whole image and local areas.The descriptor is used on the “harsh grouping” stage.Secondly,the ORB fine matching is used on “deep grouping” stage.After dividing the database into groups eliminating the near-duplicate image,the database will be simplified extremely.Thirdly,a dynamic grouping strategy is proposed by adding a “central vector” to each group.When dividing an image into one group,it only needs to be compared with the central vector which can be updated dynamically.By this way,the efficiency of the grouping algorithm is improved.Finally,an image quality assessment method is improved which considers the clarity,composition and ratio of the foreground comprehensively.When calculating the clarity,the foreground image is considered.Through this method the error evaluation of some bokeh image is corrected.The rule of Thirds Gets Trotted is introduced to evaluate image visual effect.When an image meets to the rule,the score of the image will be added.The ratio of the foreground is introduced as another rule to evaluate image visual effect.After the image quality assessment,the image with the best visual effects will be saved and the others be deleted.Experimental results show that our algorithm can simplify a database precisely and has high engineering values.