| In recent years,the space remote sensing technology has been rapid development,plenty of information resources are produced in a short period of time.Image change detection is a process in which changed regions between two images of the same scene but taken at different times are identified.It has attracted wide attention in the applications of remote sensing,medical diagnosis,and video surveillance.With the development of remote sensing technology,large parts of the earth’s surface have been observed by remote sensing sensors.Change detection in remote sensing images is of great importance,especially for disaster evaluation.Due to the fact that synthetic aperture radar(SAR)images are independent of solar illumination and atmospheric condition,they have become useful and indispensable sources of information in change detection.Recent works have paid more attention to the sparse representations for image denosing and despeckling,which have good robustness to noise and retain more effective information of the original images.And sparsity constraints are widely applied to image processing problems for extracting high dimensional features.As mentioned above,different SAR images have different level of noise and sparse feature is as a learning model of regularization in the traditional method,therefore the sparsity of representations needed are not stable and these solutions tend to be robust to or of better quality than those produced by previous models.Therefore,we use an effective MO-SFL(multiobjective sparse feature learning)model to control the sparsity of representations.We use this model as the unsupervised representation learning to obtain useful sparse representations and deal with various intensity of noise.The specific work is summarized as follows:(1)One change detection method based on multi-objective sparse feature learning model has been proposed.In view of that traditional change detection method based on deep neural network model caused instability explain and bad property by taking sparse features as the regular terms,we proposed one method based on multi-objective sparse feature learning.Specifically,this method is based on the training process of MO-SFL model.Firstly optimize the neural network layer-by-layer.Then choose one of optimized solutions to optimize next layer of network,after that fine tuning network until convergence.The features learning by this method has more discrimination which can provide better explanation of simple input.Thus it has an advantage over other methods by dealing with high-dimensional information.At the same time,according to the problem that different attributes of the remote sensing images with different noises,the model gives a more efficient solution which balance the distortion degree and sparse degree.(2)One change detection method based on multi-objective sparse feature learning model with a clustering based regularization has been proposed.In order to enhance performance of the neural network and directly use the limited training samples with low accuracy,we proposed one method based on multi-objective sparse feature learning with a clustering based regularization.It make use of a new method which construct cluster centers regular items by network features in supervised learning process of multi-objective sparse feature learning model.The network model joining clustering based regularization is k-means form that is easy to be optimized.This method not only avoids the uncertain discriminant result by sample selection but the network also can learn more complex and abstract characteristics. |
