ENTROPY BASED FEATURE FUSION IN MULTI-VIEW 3D MODEL RETRIEVAL By Zeng Pei M-B2-5507-2 Thesis Supervisor: Dr. Yang Zhixin Department of Electromechanical Engineering, Faculty of Science and Technology Abstract Virtual Reality (VR) technique is a hot issue that has been well recognized by public because of its high application value to innovative design, e-sports, online-shopping, etc. VR could construct a “real” world in a virtual environment, which is backed up with 3D modeling and manipulation technologies. Moreover, effective information retrieval from the huge available models are challenging. This phenomenon is particularly significant in product design and manufacturing industry where there are numerous 3D models being utilized. Therefore, retrieving the relevant model from massive library of 3D model has a strong practical significance to help designer to improve worker productivity and reduce development costs. Recently, retrieval of 3D models has attracted many research efforts where the similarity assessment method (SAM) is the focus of this area. Many types of SAMs have been proposed. This thesis proposes a novel Multi-view integrated method to assess 3D model similarity via combining three types of SAMs, named as Distance Shape Histogram (DSH), Solid Angle Histogram (SAH) and Isotropic Distance Shape Histogram (IsoDSH). The Particle Swarm Optimization (PSO) is applied as an optimizer to ensemble these algorithms. Entropy based method has been developed to iv define the objective function and to improve computational efficiency. Inspired by great practical success of learning with multiple distinct feature sets in machine learning, the multiple view concept is adopted to achieve a full expression of a 3D model. The feature fusion of 3D model in multi-view is superior to single view method, especially when the strengths of one view complement the weaknesses of the rest. Entropy has direct proportion relationship with the disorder degree of a system. This thesis proposes to associate PSO with entropy. The entropy with PSO are varied while searching for a solution. A low disorder degree of particles leads to a small system entropy, which may cause local optimum. The particle swarm could be used to break the local optimum by increasing the system entropy, i.e. enlarging the disorder degree. The entropy based dynamic particle swarm optimization algorithm has been developed in this study. To demonstrate the feasibility of our proposed similarity assessment method, the University of Macau Database of Shoes shapes (UMDS) which was built in IDIM lab, is adopted and organized into different categories of re-used features. In addition, the Princeton shape benchmarking (PSB) is used for the purpose of providing basic platform for further experiments. The experiment methods, including the common indicators “precision-recall diagram (P-R)”, and a new indicator “Average errors in top 9 results (E9)”, are employed. The experiment results suggest that our proposed method has excellent performance and outperforms other three single sub-algorithms.