A Generic Framework for Colour Texture Segmentation

This thesis proposes a novel method to combine the colour and the texture for colour texture segmentation. The objective of this research work is to derive a framework for colour texture segmentation and to determine the contribution of colour in colour texture analysis. The colour texture processing is based on the feature extraction from colour-textured images. The texture features were obtained from the luminance plane along with the colour features from the chrominance planes. Based on the above mentioned approach, a method was developed for colour texture segmentation. The proposed method unifies colour and texture features to solve the colour texture segmentation problem. Two of the grey scale texture analysis techniques, Local Binary Pattern (LBP) and Discrete Cosine Transform (DCT) based filter approach were extended to colour images. An unsupervised k -means clustering was used to cluster pixels in the chrominance planes. Non-parametric test was used to test the similarity between colour texture regions. An unsupervised texture segmentation method was followed to obtain the segmented image. The evaluation of the segmentation was based on the ROC curves. A quantitative estimation of colour and texture performance in segmentation was presented. The use of different colour spaces was also investigated in this study. The proposed method was tested using different mosaic and natural images obtained from VisTex and other predominant image database used in computer vision. The applications for the proposed colour texture segmentation method are, Irish Script On Screen (ISOS) images for the segmentation of the colour textured regions in the document, skin cancer images to identify the diseased area and Sediment Profile Imagery (SPI) to segment underwater images. The inclusion of colour and texture as distributions of regions provided a good discrimination of the colour and the texture. The results indicated that the incorporation of colour information enhanced the texture analysis techniques and the methodology proved effective and efficient. Abbreviations and Symbols 1D One Dimension 2D Two Dimension 3D Three Dimension CIE International Commission on Illumination CIE-XYZ Non-uniform colour space, defined by CIE CIE-LAB Uniform colour space, defined by CIE CIE-LUV Uniform colour space, defined by CIE DCT Discrete Cosine Transform GIF Graphics Interchange Format HSI Hue, Saturation, Intensity, a colour model HSV Hue, Saturation, Value, a colour model ISOS Irish Script on Screen JPEG Joint Photographic Experts Group KS Kolmogorov Smirnov LBP Local Binary Pattern LBP/C Local Binary Pattern and Contrast LVQ Learning Vector Quantisation M-KS Modified Kolmogorov Smirnov MRF Markov Random Field MRMRF Multi-resolution Markov Random Field MeasTex Measurement of Texture NLC Nearest linear Combination NN Nearest Neighbour NTSC National Television System Committee RGB Red, Green, Blue, a colour model ROC Receiver Operating Characteristic Curve SAC Symmetric auto-correlations SCOV Symmetric auto-correlations with a covariance measure SLR Single Lens Reflex SPI Sediment Profile Imagery SRAC Symmetric auto-correlations with rank order version SVR Symmetric auto-correlations with variance ratio TU Texture Unit YIQ Colour model used in American television transmission systems YUV Colour model used in European television transmission systems VisTex Vision Texture Vi Pixels in the 3× 3 neighbourhood Ei Threshold value in the 3× 3 neighbourhood uk Basis Vector EHxy Local discontinuity along horizontal direction EVxy Local discontinuity along vertical direction ECxy Local discontinuity along counter-diagonal direction EDxy Local discontinuity along diagonal direction Exy Local discontinuity measure based on the four detectors Nxy(R) Contextual neighbourhood associated with pixel (x, y) μxy(R) Mean of pixels on Nxy(R) σ xy(R) Spatial variance of pixels on Nxy(R) σ̃ xy(R) Normalised spatial variance σ max(R) Maximal spatial variance across the image σ min(R) Minimal spatial variance across the image Φ(σ̃ xy(R), θσ) Transformation based on the normalised variance θσ Threshold to limit the degree of contextual discontinuities I x,y Feature preserving adaptive smoothing I xy Intensity of the pixel (x,y) at iteration t ηij Encodes the effect of contextual discontinuities γ (t) ij Encodes the effect of local discontinuities α To determine the extent of feature preservation in terms of contextual discontinuities S To determine the preservation extent of local discontinuities x The mean of the colour plane P(i,j) pixel value at the position (i, j) G Log-likelihood pseudo metric G-statistic, similarity measure fi The frequency at bin i s, m Two sample histograms D(s, m) Discrepancy statistic or Modified Kolmogorov Smirnov statistic Fs(i), Fm(i) Sample cumulative distributions Gmax Maximum G value Gmin Minimum G value R Ratio of the maximum and minimum G value X Threshold value in splitting Smin Minimum block size MI Merger Importance value MIR Merger Importance ratio Y Threshold value in merging r a disc with radius r on the pixelwise classification d a square with a dimension d on the boundary refinement w1, w2 Weights according to the distribution of colour clustered labels MKS1, MKS2 Modified Kolmogorov Smirnov statistic for the intensity histogram and the colour histogram respectively MinMI Minimum of merger importance value kj Uniformity factor of the sample regions CLj[i] Histogram of the colour clustered labels in the sample regions Np Number of pixels in the corresponding regions e Ratio between the number of pixels incorrectly segmented by the total number of pixels in the region