Research And Application Of High-resolution Remote Sensing Image Feature Recognition Based On HLDA Model

With the launch of high resolution remote sensing satellites at home and abroad,high-resolution remote sensing images are also readily available.In the face of massive high-resolution remote sensing images,how to quickly and automatically identify ground objects has become a research focus.As the resolution of remote sensing image is constantly improved,the traditional recognition methods of "pixel-based" and "object-oriented" can only interpret the ground object categories,and cannot describe the high-level semantic information of the scene.Therefore,the hot issue of current research is the understanding of remote sensing image at the scene level.The purpose of the scene recognition task is to accurately interpret the semantic information in the scene.These semantic categories are defined by people in advance and are high-level abstract expressions of the scene content.The method based on feature coding is used to extract the underlying features of the scene for classification.Due to its good performance and ease of implementation,it has been widely used in scene recognition.How to solve the problem of "semantic gap" by connecting the low-level feature with the high-level feature is a fundamental problem that this method needs to solve.In this thesis,the related research is carried out along the path from BOVW model to theme model.Firstly,the underlying features of the scene are expressed in visual words,and then further mapped to theme expression.The main research contents and innovation points of this thesis include the following:1.The BOW model in text analysis is introduced into the remote sensing image to express the feature of the scene and realize the leap over of "semantic gap".Considering the spatial and scale characteristics of remote sensing image,the MS_BOVW model based on spatial pyramid representation is used in feature representation of scene.The experimental results show that this method is better than the classical BOVW model and the SPM_BOVW model.2.On the basis of the MS_BOVW model,the thesis proposes the h LDA model.Through the statistics of the distribution of visual words in the MS_BOVW model,it is found that the hidden theme information is the same as the cognitive results of people,which is always first roughly classified according to the feature attributes of images and then further subdivided.Therefore,this thesis introduces multi-scale features into the h LDA model,as a prior knowledge of the model hierarchy,and uses the artificial establishment of the relationship between the theme and the category to identify ground objects.3.Based on the image data of gf1-pms and the characteristics of the tea plantation,the tea plantation was extracted from the scene as the basic unit.In order to verify the effectiveness of h LDA model in tea garden identification,this study also compared the tea garden identification based on MS_BOVW model.The validity of the theme model method for tea plantation identification is verified by relevant experiments.From theexperimental results,we can see that compared with the MS_BOVW model which only uses the low-level visual features,the HLDA model can establish the relationship between the low-level visual features and the high-level semantic information,cross the "semantic gap",and get higher classification accuracy.