Font Size: a A A

Research On Local Invariant Feature Extraction Of Images And Its Application

Posted on:2013-11-28Degree:MasterType:Thesis
Country:ChinaCandidate:Y HouFull Text:PDF
GTID:2298330422973862Subject:Photogrammetry and Remote Sensing
Abstract/Summary:PDF Full Text Request
Local invariant feature extraction of images is an important area of study incomputer vision and image analysis. Local invariant features, which are invariant tomany kinds of geometric and photometric transformations and yet robust to localocclusion, play a significant role in a wide range of applications. However, there aremany requirements for variant applications and different type of images. Aiming atdeveloping local invariant feature extraction technology that is more accordant withvision perceptive characteristic, higher matching rate and stronger affine invariance,some fundamental theories and key techniques such as scale-space representation andcharacteristic scale selection, feature detection, feature description and matching aresystematically studied on the basis of technological frame of local invariant featureextraction. Furthermore, proposed methods in this thesis are applied to optical remotesensing image processing.In scale-space representation and characteristic scale selection, a multi-characteristic scales selection method and a relevant feature matching strategy areproposed, in order to resolve the problem that the distinctiveness of single characteristicscale is lower. Based on characteristic scale selection mechanism, the proposed methodsearches multiple local maximum of scale-space response, and then selectsmulti-characteristic scales to construction multi-feature regions. It contributes to thedistinctiveness of feature descriptors. Moreover, the feature matching strategy formulti-characteristic scales is designed to increase the matching rate finally.In local invariant feature detection, a new method based on Gabor filter bank,which invariant features detected has multi-characteristic scales, is proposed. First of all,keypoints are detected using2D Gabor filter bank that models the biological cognitivecomputational model. It is accordant with vision perceptive characteristic and also moreintuitional, more robust. Then, multi-characteristic scales of keypoints are selectedbased on the Gabor kernel function, which contribute to the distinctiveness of featuredescriptors. It provides strong support for improving the reliability of feature matching.Another two instructive conclusions have been reached by theoretic analysis andexperimental validation. On the one hand, it can capture salient directional changes infrequency domain and extract structural information in images well, which multi-scale2D Gabor kernel function is utilized for multi-scale analysis and multi-channel filtering.On the other hand, keypoints are detected in multi-scale analysis energy accumulatedmap. It approximates isotropy characteristics. Moreover, it overcomes the disadvantageof feature redundancy in the traditional methods.In local invariant feature description and matching, we provide an in-depth analysisof fast binary feature description and matching. On the basis of this, a new fast algorithm based on ORB, which for fully affine invariant image matching, is proposed.For one thing, it speeds up the computation of image matching that binary descriptorsare fast computed by ORB and then matched efficiently. For another, it is able toagainst large affine deformation by means of simulation intermediate images when anASIFT framework for fully affine invariant is introduced.In applications to optical remote sensing image processing, the practicability ofproposed methods is further verified by three typical applications that include automaticremote sensing image registration, automatic unmanned aerial vehicle image stitching,feature points matching of distortional images.
Keywords/Search Tags:Local Invariant Feature, Feature Detection, Feature Description, Scale-Space, Scale Selection, Gabor Filter Bank, Binary Feature, ImageMatching
PDF Full Text Request
Related items