Recognition Of Harbors And Airports On High Resolution Remote Sensing Imagery Via Saliency Analysis And Semantics

High resolution remote sensing imagery is being widely applied to fields like urban planning,agriculture census,hazard monitoring,etc.Harbors and airports are important man-made objects and they play an important role in both civil and military fields.Automatic recognition of these two objects is one of the research topics on the usage of remote sensing imagery.Harbor is a kind of complex and modular objects.It is made up of relatively simple objects including seawall,vehicle,ships,dock,buildings and roads.These objects are distributed along the coast with certain kind of spatial relationship.Similarly,airport is a kind of complex and modular objects made up of relatively simple objects including terminals,tarmacs,runways,airplanes,vehicles,buildings and various other types of accessories.These components are distributed around runway and they also have certain kind of spatial relationship.On high resolution remote sensing imagery,harbors and airports are fairly different from their surrounding ground objects.Besides,airports are different from harbors in that airports have obvious and regular line features while harbors do not because of the complexity of hydrologic condition.Traditional recognition methods fail to take these key features into full consideration and fail to take these features into full usage on high resolution remote sensing imagery.This makes the recognition results unable to meet the actual needs.This dissertation proposes a method which combines saliency analysis and semantics extraction.The method consists of the following aspects.1.According to the scale information of different components in harbors and airports,saliency analysis is applied to high resolution remote sensing imagery to generate primitive saliency maps.Then these maps are used to generate a refined saliency map.Meanwhile,fuzzy clustering and statistical methods are applied to the analysis of the refined saliency map.Based on the analysis results,regions of interest that might belong to harbors or airports are extracted from the high resolution remote sensing imagery for the object recognition.2.Based on traditional one-layer BOVW semantic model,SIFT features are extracted for feature learning,then the relational degree from fuzzy theory is adapted to calculate the image relational degree to extract semantics of higher level to build a multi-layer semantic model for preliminary and detailed classification of regions of interest that are extracted in step 1.The regions of interests that are classified as belonging to the perspective object are kept and a minimum bounding rectangle is generated to finish the preliminary recognition.3.Preliminary recognition of harbors and airports are conducted on the image which is to be identified.If the minimum bounding rectangle that shows the preliminary recognition result of harbor and the minimum bounding rectangle that shows the preliminary recognition result of airport are both generated,the refined saliency map generated in step 1 is used to extract straight line features,and a threshold is set according to the intersect or overlap situations between these straight line features and the minimum bounding rectangle that shows the preliminary recognition result of airport.Based on the threshold,a classifier is designed to avoid the confusion of harbors and airports,and to finish the final recognition of harbors and airports.Experiments are conducted on high resolution remote sensing imagery.The experiment result of the proposed method is compared with those of state-of-the-art methods to verify the effectiveness and the practicality of the proposed method.The experiment result shows that the proposed method achieves higher precision and recall rate for recognition of harbors and airports,and this proves that the proposed method can effectively identify harbors and airports from high resolution remote sensing imagery.