Research On Remote Sensing Image Recognition Based On Scene Analysis

At present,there are more and more types of remote sensing images,which have different imaging effects for different types of remote sensing images.Among them,visible light remote sensing images are the most widely used in daily life.Therefore,this thesis selects highresolution visible light remote sensing image as the research object.The purpose of the research is to explore a technical model for high-resolution visible light remote sensing image classification that can be affected by imaging time and location.The technical model proposed in this thesis can be divided into two parts:(1)Scene mining of remote sensing images.Firstly,the remote sensing image is segmented,then the low-level features are extracted,the visual words are modeled by Gaussian mixture model.Finally,the potential latent semantic analysis model is used to mine the potential scenes.This step enables grouping remote sensing images by scene.(2)Classification of remote sensing images according to feature categories.The super-pixel segmentation,sample production and underlying feature extraction are performed on the remote sensing images in each scene respectively.The classifier model in the current scene is obtained through training.The model is used to realize the classification of remote sensing images.The main research work of this thesis includes following two aspects:This thesis introduces the concept of remote sensing image scene information.A method of grouping remote sensing images according to different scene information is proposed.Firstly,the remote sensing image is evenly segmented and normalized cut,which is used to obtain visual words based on texture and color.Then,by counting the number of visual words in each remote sensing image,the word frequency vector can be obtained,and then the feature fusion can be realized.Finally,the potential scene mining of remote sensing images is carried out by probabilistic latent semantic analysis model,and the remote sensing images are grouped by scene.Compared with the traditional image scene classification method,the method of this thesis introduces the concept of visual words,so the classification accuracy is higher.Aiming at the over-segmentation of remote sensing images in traditional SLIC algorithm,a shard-merging algorithm is proposed.First,the position of the segmentation fragment is found by setting the threshold,and then the similarity of the fragment to its adjacent super-pixel tileis calculated in turn,and finally the fragment is merged into the super-pixel tile with the highest similarity.Compared with the segmentation result without fragmentation,this method removes the over-segmented fragments,which effectively reduces the number of super-pixel patches participating in the classification of features,and thus improves the classification accuracy.The effectiveness of the remote sensing image recognition algorithm based on scene analysis proposed in this thesis is verified by experiments.This thesis also designs the traditional object-oriented remote sensing image classification experiment as a comparison.The results prove that the classification accuracy of the remote sensing image classification method based on scene analysis is higher than the traditional method,and the overall accuracy reaches 93.98%.