Research On Object Recognition Based On Extreme Learning Machine

Object recognition is an important research area in image understanding and computer vision,and it has been widely used in many relevant fields.With the development of artificial intelligence and pattern recognition technology,object recognition based on machine learning has been paid more and more attention.This paper focuses on the machine learning method,aiming at researching the problem of classifier design and model construction in object recognition algorithm.The research contents of this paper are divided into two parts.First,extreme learning machine(ELM),as a new learning mechanism for single hidden layer feedforward networks,is adopted as the basic framework on object recognition in this paper to reduce the computational complexity and simplify the time-consuming training step of traditional classifiers.Second,spatial pyramid model(SPM),as a widely used object recognition model,is introduced and analyzed in this paper,and the method to overcome the high dimensionality problem of SPM by improving the construction of visual vocabulary is researched.The main creative work and research of this paper is summarized as follows:(1)The generalization ability and stability of ELM are easily affected by the randomly assigned input weights.So a new method called RFSEN-ELM is proposed.This method is formed by combination of rotation forest and genetic algorithm(GA)based selective ensemble model.In our modification,the algorithm consists of three major procedures: First,we use rotation forest to train a set of independent ELMs(RF-ELM)with high diversity.Then an optimal subset of the ensemble pool is selected by using GA based selective ensemble model.Note that GA is used here as an optimization method to reduce the negative impact of un-optimal classifiers in selective ensemble model.Finally,we use the remaining ELMs to make up the strong ensemble classifier(RFSEN-ELM).The experimental results demonstrate that the proposed algorithm shows better performance than the traditional SVM classifier and some other ensemble algorithms.(2)To overcome the major drawback of SPM due to high dimensionality of the generated feature histograms,a new model called compact spatial pyramid model(CSPM)is proposed.In this model,Agglomerative Information Bottleneck(AIB)algorithm is used to build a more compact visual vocabulary for solving the drawback of SPM,and the improvement of our method is the different fusion way of visual words.In addition,a new ensemble classifier called RFWV-ELM is proposed,this method uses a weighted voting system instead of the selective ensemble model in RFSEN-ELM.Furthermore,the condition number of matrix is added into the calculation of weights.The experimental results demonstrate that the combination of CSPM and RFWV-ELM shows better performance than the combination of SPM and traditional SVM classifier.