Research On Clustering On Texture Image

With rapid development of WWW,image retrieval,data mining and pattern recognition on large-volume image database acquire more and more attentions.As an important image feature,texture plays a critical role.Surfaces of different objects,e.g. cloud,stone and carpet etc.,exhibit different textures.Each has its own distinct texture characteristic.How to use texture efficiently can provide solid basis for detecting different images.In the meantime,as storage and network technologies grow fast, image database becomes larger and larger.People face challenges to analyze useful information from large-volume database.As one of data processing technologies, clustering is utilized to improve the speed and performance of content-based image retrieval.Clustering can also be used to improve the result of searching notably.In addition,for those who want to browse database rapidly,clustering can be used to design convenient user interface.Texture based clustering can be divided into two phases,texture feature extraction and clustering.In this paper,we studied the two phases and proposed new algorithms accordingly.The dissertation is organized in the following way.The first chapter shows the research background.In the second chapter,an algorithm on texture feature extraction using complex wavelet tranformation is proposed.The third chapter details a texture feature extraction method that combines dual tree complex wavelet and local binary pattern.In the fourth chapter,an algorithm on constructing adjacency matrix in spectral clustering is proposed.In the fifth chapter,a palmprint recognition algorithm based on major line and dual tree complex wavelet transformation is shown.The sixth chapter demonstrates a hand-shape recognition system using multiple hand fetures, including palmprint texture.Contributions of the dissertation are listed as follows:1) An image texture feature extraction and measurement algorithm based on dual tree complex wavelet and rotated complex wavelet is proposedResearch shows when human eyes see images,they will decompose images into different channels including multiple directions and frequencies,which is very similar to multiresolution analysis of signal processing.For this reason,discrete wavelet transform is adopted widely in applications because of its power of multiresolution and locality analysis.However,the image texture feature captured in wavelet domain can not present texture structure well due to lack of directional information.Dual tree complex wavelet and rotated complex wavelet developed on real DWT have the properties of multiple direction selectivity and shift invariance.The second chapter proposes an image texture feature extraction algorithm based on dual tree complex wavelet and rotated complex wavelet.Images are firstly decomposed with dual tree complex wavelet and rotated complex wavelet transformations.Afterwards,texture signature is generated from the histogram statistics information of high frequency bands.Kullback-Leibler distance is applied to mesasure the distance between the textures.2) An texture extraction algorithm that combines dual tree statistics and local binary pattern is presentedAs one of frequency domain technologies,dual tree complex wavelet transformation decomposes texture images and extracts texture information on frequency domain.However,experiements show that visually distinct texture images can be constructed by subbands with same statistics characteristics according to frequency properties of image transformations.Local binary pattern is one of space domain technologies.It extracts textures by computing the difference between the gray values of the center point and the points around center point.The third chapter presents a texture feature extraction algorithm that combines dual tree complex wavelet and local binary pattem.Firstly,dual tree complex wavelet is carried out to decompose image into subbands,from which texture signatures are then computed.Meanwhile, local binary pattern is utilized to extract texture feature on the space domain.In the final clustering step,the two features are combined to calculate the distance between two different textures.Since the two texture features complement each other well,the texture information can be efficiently exploited.3) An algorithm using dynamic factors to construct adjacency matrix is proposedSince image data normally lies in high-dimension space,dimension reduction is needed for large-volume image database before clustering.It was proven that spectral clustering can efficiently utilize the locality relationship between image data and achieve good results in dimension reduction.Spectral clustering algorithm is based on the similarity between data points and construct weighted adjacency matrix according to the simliarites.The most commonly used approach to construct weighted adjacency matrix is Gaussian kernel.However,global scale factor and fixed neighbor number make the density of data distribution difficult to be utilized.The fourth chapter proposed an algorithm to construct weighted adjacency matrix from similarity matrix using dynamic factors.For those points located in higher density,scale factor and neighbor number become larger.By this means,local data density can be detected in clustering.4) A new framework for image texture clustering is shownTraditionally,image clustering can be performed as follows.Firstly,discrete wavelet transform is used to extract texture feature.Secondly PCA or k-means is carried out on the extracted features.In this dissertation,a new framework is presented. In texture feature extraction step,dual tree complex wavelet transformation and rotated complex wavelet transformation are employed to decompose images.Then,the histogram signatures are generated from high frequency subbands.In clustering step, the adjacency matrix is computed dynamically according to the density of data distribution.On the obtained adjacency matrix,spectral clustering is used to do dimensional reduction.K-means is employed finally on the dimension-reduced data.5) A new palm recognition retrieval method on image texture is proposedAs an important biologic feature,palm can be used to determine the identity.Palm recognition is a recognition approach using palm texture information.In the fifth chapter,a novel palm recognition method is proposed.In the first step,probability distribution templates are generated based on the major lines in palms.Meanwhile, dual tree complex wavelet is used to extract detailed texture information.Then hierarchical retrieval is conducted based on the two features.Because the major line based palm recognition uses the primary information of palm texture,whereas the dual tree complex wavelet can detect the detailed information of palm texture,they can work with each other in exploiting texture information.Furthermore,the probability based major line feature and dual tree complex wavelet feature can both tolerate image rotation.The sixth chapter demonstrates an application system on palm recognition.