Scene Context Based Object Recognition In Optical Remote Sensing Images

Due to high redundancy and complex features of high resolution remote sensingimages, the objects are always lost in background, which makes it difficult to recognizeobjects and analyze scene，resulting in low recognition accuracy and high computingcomplicacy. So it is important to find a fast and effective algorithm to recognize objects inthe research area of remote sensing image. In this dissertation we propose a method ofobject recognition in remote sensing images based on scene context, because we find thatthe objects in our research are always spatially depending on scene classification, such asbridges above rivers, harbors beside sea. With the context of scene classification, we canachieve much more effective and accurate results.The main contributions in the dissertation as follows:Firstly, as the high redundancy and complex texture features of remote sensingimages, we discuss several methods of texture feature extract, and apply the LBPalgorithm to scene classification. Because of the missing problem of image intensity whenusing LBP, we improves algorithm with combining image histogram and LBP histogram,which effectively raises accuracy of classification.Secondly, with the importance of scene classification context in object recognition, amethod of scene classification based on mean-shift region segmentation and LBP featureis proposed, which focus on four scenes classification of vegetation, buildings, waters andsands. The scene context could constraint on object recognition and improves recognitionaccuracy and efficiency.Finally, we develop a process of object recognition based on scene context,depending on the spatial relationship between objects and scene context. We apply sceneclassification to this process and then, compute regions of interest and object features.After that, we adopt SVM classifier to train and classify data, and achieve the recognitionof five objects of bridges/dams, harbors, airports, road and railway terminal.Experiments indicate that the method in this dissertation has better recognitionaccuracy, and high hit rate with scene context.