Image over-segmentation, as a pre-processing step of image segmentation, splits the input image into superpixels. Those are small compact regions with irregular shapes. The majority of existing methods for texture feature extraction are not suitable for arbitrarily shaped regions. Therefore, only color information can be used to classify and merge superpixels to create final image segmentation. We propose a novel method for clustering of arbitrarily shaped image regions using orthogonal transforms with different sizes for features extraction and Gaussian Mixture Models for final clustering. Our approach is based on the properties of certain orthogonal transforms when inserting zeros into the spectrum. The proposed method is particularly suitable for classifying areas with periodic and quasiperiodic textures. However, by utilizing L∗a∗b and HSL color spaces this method can be applied to and is suitable for any type of images.