Three-dimensional Surface Texture Recognition Using RGB-D Images And Metric Learning

Three-dimensional (3D) object recognition is not only a significant branch ofcomputer vision, but also a research highlight which has long attracted plenty ofresearchers. Nowadays, this hotspot has been closely related with various kinds ofsmart products in people’s life, such as intelligent robot, intelligent video surveillancesystem and so on. Thus, the investigation of3D object recognition has importantresearch meanings and extensive application value. In the process of recognition,however, the obtained3D surface textures always suffer deformations and partialdefects, due to the different camera angles or positions. This greatly reduces theeffective information that the original surface textures can give us, and has a negativeeffect on the recognition rate directly. Therefore, looking for a method that cancorrectly identify3D surface textures is the focus of the current research on3D objectrecognition.In recent years, with the appearance of the cheap equipment which cansimultaneously capture both color and depth information of the sense, more and moreresearches based on RGB-D images have been performed to recognize3D objects.This thesis creatively uses one kind of the above equipment, such as Kinect, to obtainthe information of3D objects, and then reconstructs the3D surface textures ofobjects.In addition, this thesis introduces metric learning into3D surface texturerecognition, taking into account that distance metrics influence the efficiency ofrecognition. Our work has been done as follows: Firstly, we systematically reviewseveral common metric learning algorithms, distance metrics and classificationalgorithms. Then we use Gaussian mixture model, region growing algorithm andsome other methods to carry out a study on3D surface texture reconstruction using RGB-D images. After that, we respectively use Support vector machine (SVM), Knearest neighbor (KNN), LMNN-KNN and ITML-KNN algorithms to recognizetextures in several public libraries, such as Outex. According to comparing andanalyzing the obtained experimental results, we find that the recognition rate of ITML-KNN algorithm is better than others’. Based on this conclusion, we finally applyITML-KNN algorithm to the experiment of3D surface texture recognition, and obtainideal results.