Typical Targets Segmentation Of High Resolution Remote Sensing Image With High-precision

In the wake of developments in science and technology,the resolution of remote sensing image is getting higher and higher.High resolution remote sensing images provide important support for the recognition and classification of ground objects.On the one hand,the enhancement of the resolution makes the feature space dimension of the target increase obviously,which results in a great challenge of selecting the key feature information.On the other hand,the nonlinear mapping relationship between feature and target becomes more complex,and establishing the model is more difficult.Therefore,how to find the low rank feature space and establish the high-dimensional nonlinear mapping model of target segmentation is the key of high-precision target segmentation.After investigating the key factor of specific target segmentation with high accuracy and the state of the art in the world,this thesis focuses on the water and building segmentation in high resolution remote sensing image.The sparse feature modeling and deep neural networks are respectively employed to segment the water and building in remote sensing image,in which the advantage and disadvantage of the methodology is studied intensively.Therefore the significant technology and theory are produced for research and application of land information.The major innovative achievements of this thesis are shown as follows.(1)Aiming at water area segmentation in high resolution remote sensing image,local image entropy active contour model(LIEACM)and global image entropy active contour model(GIEACM)are proposed,both of which reduce the dependence of segmentation on the gray and improve the water area segmentation accuracy.The image entropy inside of the zero level set is adopted in CV model and forms the LIEACM.This model effectively reduces the incorrect segmentation of background where the gray value approximates to the water area with low texture complexity,which improves the water area segmentation accuracy in remote sensing image.For remote sensing image of water area with high texture complexity,the GIEACM is proposed,in which,the image entropys inside and outside of zero level set are employed in CV model simultaneously.GIEACM modifies that the level set function evolution depends on gray value,and the zero level set can evolve to the global optimal value,which improves the water area segmentation accuracy.(2)The VGG full convolution neural network architecture for building segmentation is constructed,and building in high resolution remote sensing image is segmented effectively.For building segmentation of remote sensing image,the traditional methods of segmentation are not accurate enough or even invalid.Therefore,the deep neural network(DNN)is introduced to establish the mapping between remote sensing image and the target,and the high dimension and strong nonlinearity of the mapping are described more effectively.In the VGG,the features of building are extracted automatically and the segmentation mapping is established by training.The building in high resolution remote sensing image is segmented by the proposed VGG DNN effectively.(3)An Encoder-Decoder architecture of deep learning based on ResNet is proposed,and the batch normalization is employed to improve the segmentation accuracy of building in high resolution remote sensing image.In order to improve the segmentation precision of the building,the ResNet,whose mapping is the residual error between output and input,is employed as the basic network of the deep neural network.Furthermore,the batch normalization is introduced to decrease the vanishing/exploding gradients so as to accelerate the weight convergence.Meanwhile the Encoder-Decoder architecture with multilayer feature information extracts the bulk feature and edge information of the building effectively in high resolution remote sensing image,therefore the external disturbance of building is suppressed convincingly and the building segmentation precision is improved effectively.(4)The fully-connected conditional random fields(CRF),in which the basic network is ResNet and the deep learning architecture is Encoder-Decoder,is proposed to segment the building in high resolution remote sensing image.The methodology overcomes the interruption of road whose color features are similar to the building,reduces the disturbance of staggered floor and shadow effectively.The fully connected CRF is established based on ResNet and Encoder-Decoder,in which the value of unary potential function is given by the raw result of Encoder,and the pairwise potential function states the feature of pixel-pairs in the whole image.The parameters of CRF are trained with the Encoder and Decoder simultaneously,and the segmentation results are further adjusted through the characteristic relationship between the fully connected pixel pairs.The ResNetCRF can reduce the disturbance of staggered floor and shadow effectively and obtain the best segmentation precision.In the thesis,sparse feature model and deep neural network are employed to segment the specific target in remote sensing image.The experiments and results show that there is no need of training sample in sparse feature model,but it is necessary for deep neural network.While,the generalization of the sparse feature model is weak for the complex remote sensing image to get enough segmentation precision,even fail.The deep neural network segments the specific target effectively with advanced performance of precision and generalization,however,the large number of samples,compute device and memories would be required by the deep learning.