A new approach to segment images composed of different textures is presented in this paper. We modified Texture Spectral Method, proposed by Wang et al., to decrease the number of spectral components by considering the geometrical meaning for a local change of image intensities. The number of spectral components is considerably reduced and the local strucute of image intensities can be clearly recognized by this approach. We applied the proposed approach to extract different types of image features and the experimental results on various images are presented.