Image Classification And Segmentation Algorithm Based On Clustering

Along with the rapid development of the information science and Internet technology, people can easily get a lot of image information. But because there are a large quantity of images, to analyze the image content with the naked eye becomes not reality. In order to analyze the image content automatically, we must use image segmentation and classification algorithms based on computer vision and pattern recognition technology. This thesis focuses on the research of image segmentation and classification algorithms which are based on clustering, for the image segmentation and categories can be completed with the clustering algorithm.The image representation is important in the field of image classification. A popular method is the visual words, which are constructed by k-means clustering commonly. The performance of k-means clustering severely degraded when Euclidean distance was used as the similarity measurement method because of the existence of the sparsity and noise in high-dimensional data. To solve the problem, this paper proposes a more robust similarity measurement algorithm. The algorithm combines the Euclidean distance with a similarity measurement method which is suitable for high-dimensional data. The experimental results on scene categories database and Caltech-101 database show that the performance is improved dramatically when visual dictionary is constructed through the proposed method.In order to further improve the effect of image classification, in stead of using the vector quantization, sparse coding is introduced by Yang etc to quantize the SIFT features, providing better image representation and lower quantization error rate in the process. In order to reduce the affection of the high-dimensional feature, in this paper, we apply the Principal Component Analysis to reduce the feature dimensions. The experimental results show the speed of image classification is improved.The spectral clustering algorithm is a popular research topic in image segmentation field. In this paper, we introduce a modification of“Normalized Cuts”to incorporate priors which can be used for constrained image segmentation. The researchers seek solutions which are sufficiently“correlated”with priors which allow us to use noisy top-down information and this algorithm can segment the interesting region effectively by human intervention.