Region Localization Based on Rotational lnvariant Feature and Improved Self Organized Map
The issue of target localization by means of textureanalysis is addressed First the texture featureextraction based on multi-channel Gabor filterdecomposition and the rotation invariantrepresentation of Gabor features are analyzed in theview of their ability of classification.After that,amethod based on Gabor features and neural networkclassifier is proposed The method is composed of twostages,unsupervised texture clustering and targetlocalization.In the first stage,original feature spaceextracted by Gabor filter banks is applied in training aself organized map classifier and a novel mergingscheme is presented to achieve the accuracy ofclustering.In the second stage,digital Fouriertransform of the original feature vectors are applied inback propagation(BP)network to ensure rotationinvariance in localization.In the experiments,theusefulness of the proposed method is demonstrated ontexture database and practical barcode localizationsystem as well.The method is also proved rotationinvariant and accurate in localizing target texture.
Zhuofu Bai Zhaoxuan Yang Jiapeng Wu Yang Chen
School of Electronic Information Engineering,Tianfin University
国际会议
厦门
英文
703-706
2008-11-17（万方平台首次上网日期，不代表论文的发表时间）