Researches On Object Detection In Remote Sensing Image With Complicated Scenes

Object detection in remote sensing image is critical for remote sensing applications such as area inspection, passive navigation, disaster salvation, aerocraft guiding, etc. Due to the lack of appropriate knowledge representation and reasoning methods, current object detection methods can only deal with objects of a sigle in simplifed scenes with many with many hypotheses, which restrict the scale of the problem that can be solved and the complexity of the application scenes.In this thesis, the methods of object detection in remote sensing images with complicated scenes are investigated. Promising results are attained in several aspects including the models of knowledge representation, the reasoning methods, features extraction and strategies for object searching and imaging. The contributions and achievements of the thesis can be concluded as follows:(1) A new model named object-oriented factor graph is proposed for knowledge representation and reasoning, which combines the probabilistic models with the graph models. The philosophy, methodology and related algorithms of this model are fully described in terms of basic ideas, model definitions, dynamic mechanisms, probabilistic reasoning methods, model building methods and classification abilities. Compared with existing knowledge representations and reasoning methods, the object-oriented factor graph can process expert knowledge and known information of the samples with a general probabilistic frame. And it is capable of dealing with more complicated problems with larger scale.(2) An object-oriented factor graph model, for object detection in remote sensing images with complicated scenes, is built. Firstly, the uncertainties and hierarchical characteristics of object detection in remote sensing images with complicated scenes are analyzed. Then, features for the interpretation of remote sensing images are concluded based on a thorough analysis of the process of remote sensing image interpretation by human experts. At last, an object-oriented factor graph model is built for the detection of water, harbor, and ship in remote sensing images with complicated scenes according to the domain knowledge and related features. As a result, multiple steps such as feature extracting, target searching and imaging scheduling are integrated within a general probabilistic frame.(3) Several algorithms are proposed for the feature extraction of object detection in remote sensing images with complicated scenes, including: better distinguished spectrum features in optical remote sensing images with a high resolution are proposed based on mechanisms of opposite colors; a novel algorithm for edge detection in colored images is proposed based on sector operators, which can locate edge points more accurately and is more robust compared with existing edge detection methods; an approximate computation method of the fast Gabor filter is proposed for the analysis of image textures, which needs only 49 operations on each pixel of the image regardless of the parameters and is the fastest Gabor filter computation method by now. Experiments demonstrate that its level of errors is acceptable and can deal with the analysis of textures in object detection in remote sensing images with complicated scenes.(4) A novel method of the representation and extraction of the structural features for object detection in remote sensing images is proposed based on object-oriented factor graph models together with the perceptional laws in Gestalt psychology. Firstly, common structural features for the analysis of remote sensing images are summarized and are classified into four types including key point structures, line structures, shape structures, and spatial features. Then, by reviewing the applications of perceptional grouping laws in computer vision, their potential applications in object detection in the extraction of structural features in remote sensing images with complicated scenes are proposed. At last, a novel method of representing and extracting structural features for object detection in remote sensing images is proposed based on object-oriented factor graph models.(5) A Bayesian discriminative model for object detection is built based on object-oriented factor graph models. The visual attention mechanism is introduced into the process of object searching according to several approximate reasoning methods of object-oriented factor graph, which adds to the accuracy and robustness of the detection method and decreases the computational costs of feature extraction and analysis. Firstly, biological foundations of visual attention are summarized. Then, a fast algorithm for visual attention model is proposed based on approximate Gaussian pyramid. Based on this, a target searching strategy is proposed, which is then applied to the detection of applied water, harbor and ship in optical remote sensing images with a high resolution. The results demonstrate the robustness, low costs and accuracy of the method.(6) Aiming at applications closely related to object detection such as moving target tracing and district monitoring, a beneficial new imaging scheduling algorithm is proposed for earth observation satellites based on historical information, adapting the ideas of visual attention investigated in chapter 6. Firstly, the problem of the imaging scheduling of earth observation satellites is analyzed. Then according to related historical information, the observation requests for a moving target are localized, and in the discrete observation belt, a fast computational method for the reward map of candidate imaging frames within a given time is proposed. At last, a directed acyclic graph model is built according to the map and an A* searching algorithm is proposed for the searching of the optimized imaging scheduling. The effectiveness of the method is validated by computational complexity analysis and simulation experiments.