Research On High-resolution Optical Remote Sensing Images Scene Understanding

Remote sensing image scene understanding and automatic information extraction,as the ultimate goal of the interpretation of remote sensing images,is an engineering system of remote sensing image processing field,which plays an important role in a wide applications and also has important values for both civil and military use.However,remote sensing images with complex background are often in large scale and always contain many different ground objects that vary greatly in poses,size and orientation.In addition,remote sensing images also contain variety of image scenes with similar appearance that can be easily confused with each other.These factors have brought about many difficulties for scene understanding and automatic information extraction of remote sensing images.Consequently,how to automatically understand the remote sensing image and extract important information from remote sensig images is a meaningful and challenging research topic.This dissertation studyes high-resolution optical remote sensing images and performs researches on the key technologies of remote sensing image scene undertanding such as object detection,scene classification and semantic annotation.To overcome the limits of the existing methods,we propose four novel methods for these subtasks of image scene undertanding.As the research focus of remote sensing image processing,the presented studies have important theoretical and practical value.The main work and contributions of this dissertation are summarized as follows:1.In order to effectively address the problem of detecting objects from large-scale remote sensing images,a novel coase-to-fine approach is proposed by using visual saliency and condition random filed(CRF).It works in a hierarchical architecture with a coarse layer and a fine layer.At the coarse layer,a target-oriented saliency model is built by combing the cues of contrast and line density to rapidly localize the candicate region of interest(ROI).With coarse layer,on one hand the target search region can be reduced significantly,which will result in a lower false alarm rate.On the other hand,the computational cost can also be reduced and the detection efficiency is improved accordingly.Furthermore,at the fine layer,a learned CRF model is applied to each ROI to perform the fine detection of the desiring objects.The CRF model is learned based on sparse features of local patches in a multi-scale structure and it also takes the contextual information of target into consideration.Therefore,its detection is more accurate and is robust to target scale variation.Comprehensive evaluations on RSI database from the Google Earth and comparisons with state-of-the-art approaches demonstrate the effectiveness of the proposed method.2.In order to effectively address the rotation problem caused by the imaging perspective(i.e.from top to down imaging)and the semantic gap problem caused by the low-level visual features,a novel rotation-invariant multi-feature probabilistic Latent Semantic Analysis(RIMFp LSA)is proposed for both object detection and scene classification.Firstly,different kinds of low-level image features are extracted and quantized to obtain the visual words and visual dictionary,respectively.Then,the two different visual words are jointly modeled with their conditional distributions constrained via a single latent semantic topic,which make us to take full advantage of the complementarity between the different features.Meanwhile,a rotation-invaritant regularization term is applied to enforce the leanred latent topic from different rotation directions of the same image to be as consitent as possible.Thus,the learned topic can deal with the rotation problem of remtoe senining image processing.Comprehensive experiments about object detection and scene classification demostrate the efficitvess of the proposed method.3.In order to learn discrimative high-level features from remote sensing images for scene classifcation,a novel stacked discrimative sparse autoencoder(SDSAE)is proposed to perform sence classification.Firstly,a novel discrimative sparse autoencoder(DSAE)is desinged by using pairwise constraints to servse as the discrimiative regulation term.The pairwise constraints include must-link constraints and cannot-link constraints,where the former indicate the similarity of two images with the same semantic class and the latter indicate the similarity of two images with different classes.The pairwise constraints guarantee the ability of discriminative representation by enforcing the images with the same class to be close in the learned feature space while the images with different classes to be kept far away.Then,the DSAE is stacked to form a novel deep neural networks(i.e.,SDSAE)to learn more discrimative high-level features.Finally,a softmax layer is constructed on the top of SDSAE model to perform scene classification.Comprehensive experiments on LULC dataset demostrate the efficitvess of the proposed method.4.In order to achieve better automatic image semantic anntotaion performance with limited training samples and minimal labor cost,a novel high-level feature transferring annotation method is proposed.The proposd method exploited the transferred high-level features to serve as the representations of remote sensing images instead of learning high-level features from trianing data at hand,which will reduce the demand of the training data while while ensuring the robustness of image representations.We also use a weakly supevsied manner to learn a multiclass classifier which differs from the existing fully supervised method that needs to prepare large amount of training samples with high-quality pixel-level labels.The weakly supevised manner largely reduce the diffculty of labeling training samples and make the method more suitable for current applications.Comprehensive experiments on Sydney and Washington images demostrate the efficitvess of the proposed method.