Research On Scene Classification Of Remote Sensing Images Using Optimization Of Visual Vocabularies Concerning Scene Labels

With the rapid development of remote sensing technology,the sensor technology has become more and more mature,and the spatial resolution of remote sensing image is getting higher and higher.How to classify land use and land cover categories in high resolution remote sensing image is one of the hotspots of remote sensing image classification.Traditional pixel-based or object-based methods have been unable to solve this problem perfectly,so we need to dig out the deeper abstract features of remote sensing images to better express these land cover and use scene types.Semantic features are an abstract feature that is often used in remote sensing image classification.The Bag of words model is a traditional semantic feature.However,the visual vocabularies of the Bag of words model usually does not consider the information of scene labels,So we need to develop a common visual word optimization method,the innovation and the main research content are as follows:In this paper,we propose a visual word optimization method which takes into account the information of scene labels,and generates a visual dictionary for each category,so that each category has its own unique visual vocabularies,enhancing the recognition ability of visual vocabularies.The experimental results show that compared with other methods that do not use the visual dictionary for each category,the proposed method makes significant improvement in terms of classification accuracy.The extracted features in the feature extraction step are not always useful for scene classification.These useless features will increase the computational complexity during the vector quantization process and affect the representation of the image.Therefore,the extracted features need to be selected.This paper presents an improved iterative keypoint selection method to remove those keypoints that are redundant or keypoints that appear in most scenes.Experiments show that this method can not only improve the efficiency of the calculation,but also improve the classification accuracy.On the basis of the visual dictionary for each category,this method uses a two-step classification method to further improve the classification accuracy.In the method incorporating the information of scene labels,the generated BOW features are typically mapped to each category to obtain a histogram for classification in each category.If the generated BOW features are mapped to the visual dictionary in the inappropriate category,the predicting result may be incorrect,and the histogram of this particular category will have a negative effect on the output of the classification.Therefore,we first use the KCRC classifier to get two candidate categories that are most likely to be true categories,and use the class-specific histograms of the two candidate categories to be put into SVM.The experimental results show that the accuracy of classification improves obviously using KCRC classifier.