Research On Point Detector And Descriptor For Textureless Object Surface Based On Wide Baseline Stereo

Digital photogrammetry has become an important way to obtain3D spatialinformation for its benefit of quick, flexible and untouched styles. Since manyproblems related to this field have been proposed, extracting corresponding pointsbetween images from different viewpoints is one of the most important problems needto be solved among them. During the past years, it has been taken the growing interestin lots of researchers, almost methods of solving this problem are based on thetextured image regions with intensity variations in scale and space, since extractingpoints from these regions have distinctive descriptions which are easy to be matched.However, broad categories of real-world objects have non-textured regions. Regulardetectors are very likely miss stable of feature locations there. Additionally, mostkinds of descriptors are built using local image gradients and thus lose theirdistinctiveness when built on non-textured areas. Hence, non-textured regions usuallyare not considered for detection and description.The research work of this dissertation mainly focus on point correspondingbetween wide-baseline images with weakly texture and involved following methods,such as pattern recognition, combinatorial optimization algorithm and digital imagingprocessing mainly include:(1) The basic method of point detection, description and stereo matching areintroduced, a detailed exposition and summarization in these areas are provided, suchas sparse representation, graph cuts and SIFT, which are used in the followingchapters.(2) For the areas of low textured in image pairs, there is nearly no point that canbe detected by traditional methods. The information in these areas will not beextracted by classical interest-point detectors. A novel weakly textured point detectionmethod is presented. The points with weakly textured characteristic are detected bythe symmetry concept. The proposed approach considers the gray variability of theweakly textured local regions. The detection mechanism can be separated into threesteps: region-similarity computation, candidate point searching，and refinement ofweakly textured point sets. The mechanism of radius scale selection and texture strength conception are used in the second step and the third step, respectively. Thematching algorithm based on sparse representation (SRM) is used for matching thedetected points in different images. The results obtained on image sets with differentobjects show high robustness of the method to background and intraclass variations aswell as to different photo metric and geometric transformations; the points detected bythis method are also the complement of points detected by classical detectors from theliterature. And we also verify the eiffcacy of SRM by comparing with classicalalgorithms under the occlusion and corruption situations for matching the weaklytextured points. Experiments demonstrate the effectiveness of the proposed weaklytextured point detection algorithm.(3)Since the image feature points are always gather at the range with significantintensity change, such as textured portions or edges of an image, which can bedetected by the state-of-the-ml intensity based point-detectors. There is nearly nopoint in the areas of low textured detected by classical interest point detectors. So wedescribe a novel algorithm based on aiffne transform and graph cuts for interest pointdetecting and matching from wide baseline image pairs with weakly texture object.The detection and matching mechanism can be separated into three steps: firstly, theinformation on the large texlureless areas will be enhanced by adding textures throughthe proposed texture synthesis algorithm TSIQ (Texture Synthesis based on ImageQuilting), we use the sn^e-based scheme to detect the contour of the large texturelessobject, on which the feature points close to the object contour will be used to estimatethe aiffne transform between image pairs. Secondly, the initial interest-point set isdetected by classical interest-point detectors. Finally, graph cuts are used to find theglobally optimal set of matching points on stereo pairs. The efficacy of the proposedalgorithm is verified by three kinds of experiments, that is, the influence of pointdetecting from synthetic texture with different texture sample，the stability under thedifferent geometric transformations and the performance to improve the quasi-densematching algorithm, respectively.(4)Though dense short-baseline stereo matching is well understood，itswide-baseline counterp^is, in contrast, much more challenging due to largeperspective distortions and increased occluded areas. It is nevertheless worthaddressing because it can yield more accurate depth estimates while requiring fewerimages to reconstruct a complete scene. On the other hand, local region detectors canbe used to get robust corresponding points under wide baseline image pairs. However,the traditional descriptions can not be used for points in the weakly texture regions. We therefore introduce a new method to descript points in these regions, which usesregular interest points to descript them. Moreover, we use this description to get widebaseline stereo matching, we demonstrate that our approach is not only suitable forthe description of non-textured areas but that precision is significantly superior tothose of regular features.Our research work in this paper provides some new methods concerned on pointextraction and corresponding with wide-baseline images, which can be used in stereomatching with two uncalibrated images.