The Image Classification Algorithm Based On Feature Extraction And Sparse Representation

With the development of computer science, computer vision has become an emerging discipline. The last decades, computer vision gradually integrates in the areas of machine learning and artificial intelligence. This paper is main research of remote sensing image classification and identification of issues mainly by aviation remote sensing equipped with sensors or high resolution satellite imaging to a particular geographical area. In space level the upgrade sensors keep high resolution remote sensing image direction covering from a pixel-level of only a few meters high-image, making remote sensing image-class collection of high quality, which are crucial to computer vision and pattern recognition. The main contributions of the paper are:First, to overcome the issue that in hyperspectral image classification in extracting features by the fixed image widow when on the edge region will introduce irrelevant features, we proposed an adaptive spatial spectral based on support vector machine feature fusion algorithm for hyperspectral image classification, which utilizes the unsupervised algorithms to generate the contexture information and makes fusion of contexture features for SVM.Second, to overcome the issue that in hyperspectral image classification in joint sparse coding using a fixed window of signals acquisition in the image area edge will easily introduce impurities, an adaptive window based on sparse coding space spectral characteristics of combined reconstruction algorithm for hyperspectral image classification is proposed, which utilizes the unsupervised contexture information to sparse code by joint SOMP, and uses the reconstruction errors for classification.Third, to overcome the issue that in hyperspectral image classification the drawbacks of k nearest neighbor algorithm using Euclidean distance as similarity measurement between high dimensional features, a nearest neighbor classification based on sparse coding algorithm for hyperspectral image classification is proposed, which utilizes the sparse coding coefficient instead of Euclidean distance similarity, and further generalizes to spatial space to generate classification rules.Finally, to overcome the issue that different spatial scales features can not be effectively fused in scene classificastion, we proposed a global feature in conjunction with local feature reconstruction fusion of natural scene categorization algorithm, which utilizes the fusion of multi-features for scene classification.