Research On SIFT Descriptors Selection Technology For Image Retrieval

Image feature representation is a basic research area in computer vision,and it plays a fundamental role in the study of many applications.One of the most classic and widely used feature descriptor is SIFT(Scale Invariant Feature Transform)which was proposed by David Lowe in 1999.SIFT is famous for its robust scale and rotation invariance,and is widely used in image retrieval,image stitching,3D reconstruction and other fields.However,with the coming of big data multimedia,the application scenarios become more and more complex,making the conventional SIFT features face many problems: The large number and high dimension lead to SIFT need huge storage space and low calculation efficiency in application.Recently,the improved algorithms of SIFT features emerge in endlessly,and can be divided into three categories: Firstly,dimensionality reduction,such as the famous SURF(Speeded Up Robust Features)and PCA-SIFT(Principal Components Analysis-based representation for SIFT),which have achieved good performance in many applications.Secondly,number reduction.It has emerged some methods in this category recently,but the overall development is not mature and many of them have big limitations in scalability.Lastly,improve the approximate solution such as hierarchical structure and LSH(Locality-Sensitive Hashing).The study in this paper mainly belongs to the second one,we reviewed the background of SIFT points selection and investigated the existing methods sufficiently.Through analysis and comparison among these methods,we found many advantages and limitations of the recent algorithms.Then we propose a new SIFT points selection method based on dictionary learning,and apply it into the area of image retrieval to demonstrate its effectiveness.The primary research in this paper aims to analyze the correlation between SIFT descriptors and remove redundant ones.Then we choose a small number but more powerful feature set.During the study we found there exists strong inner relativity between SIFT points selection and dictionary learning for sparse representation,then we turn our problem into dictionary learning.We design a new dictionary learning method to adapt our problem and employ simulated annealing algorithm to obtain the optimal solution.During the process of learning we combine the sparsity constraint and spatial distribution characteristic of SIFT points.And lastly select small representative feature set with good spatial distribution.The experiments verify that the selected SIFT points have good performance,which can not only save storage space and improve time efficiency,but also maintain the accuracy in applications.This method is unsupervised,which does not rely on any dataset,so it has stronger scalability.