Feature Extraction In Light Field Based 3D Object Retrieval

Producing and displaying of 3D objects are conventionally based on models. However, 3D model building is of high complexity. In order to solve this problem, light field based 3D object description has been innovated recently, i.e. to describe 3D objects through dense multi views with different view points. Therefore 3D objects are transformed into a series of 2D images. Relevant retrieval mechanism has been developed subsequently. The main bottleneck of light field based 3D object retrieval includes information redundancy, spatial correlation and isomerism of multi views, hence, techniques on 2D feature extraction can not meet the demand of retrieval here. To alleviate this constraint, this thesis mainly focuses on the feature extraction of light field based 3D objects.Firstly, an algorithm is proposed for extracting characteristic views, i.e. selecting views which contain maximum amount of information from large number of light field 3D object multi views. Based on this, the thesis develops an improved spatial distribution descriptor, to address the limitation of SIFT feature descriptor which can only cover local region information around feature points. This method fully utilizes global and local information to improve retrieval efficiency and accuracy, while reduces number of views for describing a 3D object.Secondly, the thesis proposes a multi-feature fusion methodology. By fusing Shock Graph descriptor and Fourier descriptor, this methodology combines advantages of these two shape descriptors while makes up for their own limitations, thus improves retrieval effectiveness furthermore.Besides, an approach of feature extraction based on image segmentation is proposed. The feature descriptors are respectively extracted from segmented regions of the whole view. This approach integrates spatial correlation of multi segmented regions. Experiment results showed the superiority of this feature descriptor for light field based 3D object.