Study On Boltzmann Machines For Robust Target Recognition

Target recognition is an important research area in computer vision,and the core part of many civil and military applications.In recent years,"deep learning" based target recognition method has become a great success under the context of big data.Compared with traditional methods of computer vision,deep learning utilizes large dataset more efficiently,and can even lead to some high level representations.However,this excessively relying on data can cause some potential problems.Firstly,deep learning excessively relies on supervising signals,which makes it sensitive to noise or perturbation,and more likely to overfit the training data.Secondly,deep learning cannot efficiently represent ambiguity.Focusing on these two problems,this paper aims at proposing robust target recognition method suitable for relatively smaller dataset.The basic model investigated in this paper is the infinite restricted Boltzmann machine,which is capable of unsupervised learning and adaptive model complexity selection.Meanwhile it is very flexible on the model structure.These properties make it more suitable for small datasets.Firstly,we propose the discriminative infinite Boltzmann machine to make the original method capable of supervised learning and semi-supervised learning.The proposed model can model the joint distribution of data and label.It can also represent the posterior distribution of label conditioned on data.However,due to the "ordering effect",the convergence of learning is slow,and it is easily to result in redundant hidden units.In order to deal with the slow convergence of learning,we propose a new optimizing method for the model.The core idea of this learning method is firstly replacing the assumption that the order of the hidden units is fixed by random order of hidden units.Random ordering the hidden units can greatly eliminate the dependency between hidden units.Further analysis indicates that the learning method can make the model invariant to the permutation of hidden units.Experimental results also validate that the new training method can result in faster convergence and better generalization.This algorithm is also the core and basic learning algorithm for all the models studied in this thesis.After that,based the new training strategy,we apply the infinite restricted Boltzmann machine on multi-view radar high resolution range profiles based target recognition.The contrastive experiments with the hidden Markov model and the restricted Boltzmann machine shows that,the proposed model learns the features more effectively than the hidden Markov model and can adjust its size more adaptively than the restricted Boltzmann machine.By utilizing the generative property of the proposed model,we also apply it on the tasks in which range profiles from some views are missing.The model can infer the missing range profiles from the observed range profiles.And the completed range profiles are used for recognition,which greatly alleviates the influence caused by the missing rage profiles.In order to further increase the representational power of the model,we propose a multi-layer infinite Boltzmann machine,and refer to it as the infinite deep Boltzmann machine.The number of hidden units of each layer of the proposed model is changeable,which makes it more adaptive and can greatly release the burden of manual intervention.Meanwhile,in order to deal with continuous valued data,we propose the infinite Gaussian restricted Boltzmann machine.The model can adaptively learn the mean and standard deviation with the hidden units.By combining these two models,the variable type we are dealing with can be greatly expanded.Experimental results indicate that the model can achieve the performance competitive to classic deep Boltzmann machines with more compact model size.In order to better utilize the high dimensional long sequences,we propose the infinite conditional restricted Boltzmann machine,along with a multi-scale structure for representing long range dependencies in the sequence.The infinite conditional restricted Boltzmann machine can represent conditional distributions of high dimensional sequences,thus can be used for prediction and classification of sequences.Meanwhile,it is also capable of adaptively learning the model size.We apply the model on sequential range profiles recognition,and the model can achieve high recognition rate.We also apply the model on interrupted sequential range profiles recognition.By utilizing the model to predict the range profile at the next time step based on previous sequence,the sequence can be completed,which greatly increase the robustness of the recognition process.Finally,we propose two methods of incorporating prior knowledge into the model.The first is structuralizing the connections between hidden and visible units,which leads to the localized infinite restricted Boltzmann machine.The localized connections can encourage the model to better learning local structures in the image.The model can generate different local receptive field automatically during training.Experimental results indicate that,the localized connections can achieve better generalization especially when the training dataset is small.The second method is the sparse activation constraint for infinite restricted Boltzmann machines inspired by sparse representation.The constraint can effectively control the training process of the model,and makes the hidden units more selective.Experimental results on various datasets also validate that the proposed constraint can successfully guide the model to be more sufficiently trained.