Research Of Image Retrieval Based On Multiscale Geometric Analysis And LBP

With the rapid development of digital multimedia techniques and Internet, image has already played an important role in the multimedia information system and widely employed in many important applications such as:digitallibraries, medical image man-agement, remote sensing image processing and other fields. How to retrieve the required image quickly and efficiently from large image database is a significant and serious chal-lenging research topic. At present content based image retrieval has become a hot research. Feature extraction is critical to content based image retrieval, and texture is one of the basic visual features.Multiscale geometric analysis comes from and higher than wavelet. It is a new image analysis tool, which is multi-directional and anisotropic. And it could extract image fea-tures such as texture, edge and other features more effectively. Local binary pattern (LBP) is one effective texture descriptor to represent texture images. Therefore, this paper com-bines multiscale geometric analysis with LBP, and carries out three parts study on image retrieval.1.This paper proposes a texture image retrieval method via LBP feature and Brushlet domain coefficient statistic feature. More specifically, frequency feature is described as energy feature of every Brushlet subband. Spatial feature is LBP histogram extracted from images. Similarity between images is measured by improved Canberra distance. Further, the suitable weight is chosen for image retrieval.2.This paper proposes an adaptive image retrieval method via spatial-frequency mixed features. This method can describe spatial and frequency information of image simultaneously. More specifically, spatial feature is LBP histogram extracted from im-ages. Frequency features are described as the generalized Gaussian density (GGD) of Contourlet transform detail coefficients and mean value and standard deviation of ap-proximation coefficients. Further, we use closed-loop feedback to adjust weighting factor adaptively for image retrieval. Experiments show that average recall rate of this method is 10.4%,8.0% higher than frequency domain method and LBP respectively.3.This paper proposes a novel rotation invariant texture descriptor CLBP-HF. The histograms of uniform pattern are extracted from images, and then Fourier power spec- trum is obtained to form the global rotation invariant feature. The dimension of feature is reduced because of the symmetry of feature. Finally improved Canberra distance is adopted for image retrieval. Experimental results show that based on rotation texture database CLBP-HF can achieve significant improvement, compared to traditional rotation invariant LBP.