Reflection Component Separation And Spectral Compression Of Texture Images

Color texture image consists of two components, i.e., color and texture, and the color is considerably influenced by texture. Due to the wide application of color texture image, as well as its own particularity, color texture image processing has been intensively studied in the literature. In this thesis, we mainly propose some innovative algorithms in the separation of reflection components and the spectral compression which are of theoretical and application value.The study of reflection component separate is to decompose the reflectance (color) and shading (texture) components from an image. In the algorithm, we first use the seed region growing technique to accomplish image segmentation, and then extract the initial texture component based on the reflection model. Third, we use the texture synthesis technique to restore the regional boundaries. We then obtain the reflectance component from the reflection model, and refine the reflectance by non-local denoising.Multispectral images can provide much more spectral information than 3-channel color images. Due to the huge data volume, it is necessary to compress multispectral images for efficient transmission via intranet/internet. The adaptive spectral compression method proposed in this thesis is based with the correlation between the different spectral channels. Multispectral image is initially segmented by the hierarchical analysis of the projection coefficients in the global subspace, and then ambiguous pixels are identified and classified into proper clusters based on linear discriminant analysis. The dimensionality of each adaptive subspace is determined by specified reconstruction error level, with further cluster splitting if necessary. The efficiency of the proposed method is verified by experiments on real multispectral images.