A Study Of 3D Model Classification Based On Extreme Learning Machine

Extreme Learning Machine has the advantage of a better algorithm structure,faster training speed,good generalization performance and do not fall into local optimum.Therefore,it has been widely used.3D model data has the problems of high dimension,small samples,high noise and class imbalance.In order to solve the problems of 3D model data mentioned above and the unstable performance of Extreme Learning Machine,this thesis focus on Extreme Learning Machine algorithm to do further research on 3D model data.The golden optimization algorithms and restrictions Boltzmann machine algorithm are combined with Extreme Learning Machine to improve the classification accuracy on 3D model data,and the idea of transfer learning is introduced to deal with the class imbalance problem of 3D model data.The main innovation points are as follows:(1)An improved algorithm of Extreme Learning Machine based on golden section optimization algorithm is presented.Golden section optimization algorithm is used to optimize the number of hidden layer nodes,the weights of input layer and the deviation of hidden layer.The experimental results show that compared with the traditional BP neural network,support vector machine and extreme learning machine,GS-ELM algorithm can achieve higher classification accuracy.(2)An improved algorithm of extreme learning machine based on restricted boltzmann machine(RBM-ELM)is presented.Restricted boltzmann machine is used to optimize the weights of input layer and the bias of hidden layer,meanwhile to extract discriminative low-dimensional features from the raw data.The experimental results show that compared with the random forest,logistic regression,support vector machine and extreme learning machine,RBM-ELM algorithm can achieve higher classification accuracy.(3)An improved transfer learning algorithm based on Extreme Learning Machine is proposed.This method is aimed to use the correlation between plenty of labeled source domain data and target domain data with few samples to find out the related samples,improved the transfer learning's performance on imbalance data.Experimental results on the Princeton 3D Shape Benchmark Database show the algorithm has better classification results and good stability.