Researches On 3D Model Classification Based On Neural Network Selective Ensemble

3D model is a natural and direct way to illustrate the real-world objects. With the proliferation of 3D models, it becomes an emergency task to obtain the desired models from the existing. This topic can facilitate the modeling process, and has great value in manufacturing, military, virtual reality, simulation etc. Therefore, 3D model retrieval emerges as an important field in multimedia retrieval, which aims at retrieving the desired models correctly, quickly and conveniently.The context-based 3D model retrieval is research focus at present due to the drawbacks of text-based retrieval. Nowadays, the content-based retrieval technique has many improvements in the theory and the applications. The meaning of content-based is to use 3D model visual features to reflect information of content, automatically set up the feature-index. First, automatically calculate and extract the features of 3D models from the model data, such as shape, spatial relationship, materials color and texture, set up the multi-dimensional information-index of 3D models, calculate the degree of similarity between query model and the target model in the multi-dimensional feature space, to achieve the purpose of retrieval.Classification is an important research topic in the data mining, machine learning and pattern recognition field. Classification techniques apply to handle the large high-dimensional database, such as voice, image, video and other multimedia data. The purpose of classification is based on the features of database, rely on pre-defined categories of training examples, construct a classification function or model, this structure can map a sample of with unknown category to a certain category. For the huge amounts of data, excellent classification algorithm can avoid retrieval efficiency low due to excessive and unrelated compare, allow the scope of retrieval reduced, and improve retrieval efficiency and effectiveness of results.The thesis conducts researches in the key technology in 3D model classification, use methods and theory of machine learning, construct a architecture of classification for 3D model, propose a better classification algorithms, play a positive role in organizing the 3D model database.Optimal design of classifier is the key step of 3D model classification. The main methods of classifier constructing are decision trees, neural networks, Na?ve Bayes, k-NN kernels,rules learners etc. Neural network can makes the output akin to target by adjusting the weight coefficient, also has good fault-tolerant capability, classified capability, parallel processing capability and autonomous learning ability. A better neural networks have the following characteristics:(1) Moderate complexity of the network structure;(2) Contain samples with amount of information;(3) Have features with greatest contribution to classification rules;(4) Higher accuracy of classification.Select BP neural network as classifier. BP neural network is a multi-layer neural network, training according to supervised learning approach. Provide the feature vector and the category vector to the neural network, neuron activation values transmit from the input layer via the hidden layer to output layer, With the principle of reducing the error between the desired output and actual output, amend the connection weights layer-by-layer reversely. With the training process ongoing, the accuracy of neural network will continue to improve in response to the input mode. In this thesis, classified the model in 3D model database based on BP neural network, can achieve model-classification more accurately, verify the applicability of BP neural network in 3D model classification, with a certain degree of accuracy.However, a single neural network exist the questions such as difficult to convergence; the classification accuracy is not high, in the thesis, the idea of"Selective Ensemble"in the field of machine learning is applied to 3D model classification. Selective Ensemble BP neural network classifier, the result is decided by integrating various neural network outputs. Bagging algorithm is applied to the process of generating individual neural network with difference; reduce the generalization error of ensemble, lay the foundation for selective ensemble.How to realize selective ensemble BP neural networks? Classification algorithm design relies on two aspects: the degree of difference and local validity. Differences have a great influence on the generalization error for selective ensemble neural networks, the greater of differences, and the stronger of generalization after selective ensemble. In this thesis, select better neural networks to constitute a set by comparing the differences between individuals. On this basis, use local validity in the 3D models sample-space, select neural networks that have perfect results on similar 3D models as a secondary set, optimize the Classification algorithm. The experiment shows that this selective ensemble neural networks algorithm has achieved better recognition result than a single neural network, Bagging-based ensemble, K-Nearest Neighbor ensemble and other algorithms, achieve a purpose of effective classification.At the same time, classification algorithm has played a definite positive effect on classification and re-organization of the 3D model database. Effective classification of 3D models is a basic search topic in 3D model retrieval, some problems such as evaluation of retrieval results, comparing the merits of the features, feature combination, classification information are necessary. Existing researches pay more attention to the effects of retrieval, lack of enough concern on organization of 3D model database. Absence of efficient organization of massive 3D models, will inevitably affect the practicality of retrieval. The classification algorithm in this thesis can classify and recognize 3D models quickly and accurately, re-build a new organizational structure, provide better service for 3D model retrieval.On the whole, the thesis pursues the research of classification topic in 3D model retrieval, and particularly concentrates on optimal design of classifiers and selective ensemble classifiers. The classification algorithm in the thesis provides a new approach to classify the 3D models. It also plays a positive role on organization of the 3D model database. Re-build a new structure of the database through the classification algorithm, can better serve the 3D model retrieval.In future works, the proposed methods in this thesis will be integrated in the efficiency and applicable 3D model retrieval system, in addition to perform in-depth researches based on current work.