Research On Local Image Features Based On Scale-invariant Feature Transform

Image registration is the foundation of computer vision and other techniques.Because of the invariance against many kinds of transformation and robustnessagainst cases such as overlapping and sheltering, local image feature technique hasbeen a hot research direction. Among all kinds of local image features, Scale-InvariantFeature Transform (SIFT) shows good robustness against rotating and scalingtransformation, and therefore has been widely paid attention to, researched andapplied. However, SIFT is still to be improved on both matching accuracy and speed.This paper introduces the components and development of local image featuretechnique and some of the most popular local image feature methods including SIFT.Based on the detailed analysis on SIFT, this paper proposes four improvements onSIFT, including:(1) algorithm based on adaptive distance ratio matching via matchingcredits;(2) feature description via weighting by variance and gradient direction ofdescriptor components;(3) multi-layer matching algorithm via source imagesegmentation and retrieval;(4) hash-based speeding up algorithm on feature detectingand matching. This paper also combines the four improvements mentioned above andproposes two more comprehensive methods respectively aiming at the improvementof matching accuracy and speed. For each of the improvement mentioned above, thispaper shows the statistics and analysis on experiment results and comparison withsome existing methods.The main contributions of this paper are listed as follows:1. This paper introduces some of the most popular techniques on local imagefeature indetail and compares them with each other and proposed methods throughexperiment results.2. To improve the matching accuracy of SIFT method, this paper proposes threemethods: algorithm based on adaptive distance ratio matching via matching credits,feature description via weighting by variance and gradient direction of descriptorcomponents and multi-layer matching algorithm via source image segmentation and retrieval. Moreover, a comprehensive method composed of above three methods isproposed;3. To improve the speed of SIFT method, this paper proposes hash-basedspeeding up algorithm on feature detecting and matching. Moreover, combined withfeature description via weighting by variance and gradient direction of descriptorcomponents, this paper also proposes a comprehensive method to improve the speedof SIFT method.