Multi-spectral Image Registration And Recognition Based On Local Features

In this thesis, we discuss and research multi-spectral image registrationand recognition based on local features. Multi-spectral images come fromdifferent imaging mechanism with various imaging characteristics, and thegray levels differ significantly between images. Traditional image processingmethods based on gray levels often represent global features only. Since globalfeatures rely on image gray distribution and the diversity of multi-spectralimage gray levels, global methods often ends up with bad results.We deal with this situation by using local features which have been deeplystudied and widely applied in many areas of computer vision. First, weevaluate SIFT feature in the match of infrared and visible images, and then wepropose a new edge-enhanced MSER (E-MSER) feature detector combinedwith SIFT feature descriptor to process the multi-spectral image registration.Results show that SIFT feature performs very well in the infrared-visible imagematch test. As for the more complicate situation, the proposed method E-MSER succeeds in extract region features and achieves high registrationaccuracy. Comparison experiments also show E-MSER an advantage overother detectors with error less than1pixel. Results under scale and affinetransformation prove that E-MSER method is robust to this imagetransformation.For the part of image recognition, we absorb the idea of Bag-of-Words(BoW) to represent images abstractly. A dense sample is taken to extract SIFTfeatures at first, and the codebook is built by clustering these features using K-means method. Then we can get a BoW representation of the image and feedit into a SVM classifier. Since the BoW method drops global distribution whichis very useful in recognition tasks, we introduce a spatial pyramid matching(SPM) method to the BoW framework to make up the spatial distribution.Sparse encoding methods is also introduced to accelerate the speed of vector operation. Our experimental results show the classification accuracy hasreached the state-of-art. Besides, we conduct a research on the influence to theclassification accuracy of background images.