Color texture segmentation using quaternion-wavelet filters and SOM

This paper proposes a novel framework for color texture segmentation based on Discrete Reduced Biquaternion Fourier Transform (DRBFT) and Discrete Wavelet Transform (DWT). Reduced biquaternions (RB), which are an extension of the complex numbers, are used to characterize the color information. Multi-Channel wavelet filtering is used to extract the texture information at various scales from the LUV color space. Each wavelet channel (subband) is represented by a commutative Reduced Biquaternion representation. These color and texture descriptors are merged in the frequency domain in-order to reduce the time complexity of this convolution process. Amplitude features are computed from each color-texture channel. Every pixel in the image is described by a feature vector, formed by grouping the features of all the channels. Finally, SOM (Self-organizing Map) with adaptive learning rate is used to label the pixels. The proposed segmentation algorithm is experimented on some natural images from Berkeley segmentation database. The obtained results substantiate the feasibility of the method, whose performance is compared, for evaluation, with the popular JSEG algorithm and also with human-assisted segmentations.