Table 2.

Approaches proposed by other authors with which the results obtained by the proposed approach were compared.

Colour-Texture Segmentation Algorithm	Summary
Hoang et al. [98]	The colour and texture information is included in the segmentation process. The RGB image is converted into a Gaussian colour model. Primary colour–texture features are extracted from each colour channel using a set of Gabor filters. Feature vectors, whose dimensionality is reduced by applying Principal Component Analysis, and used as inputs for a K-Means algorithm, providing initial segmentation that is refined by a region-merging procedure.
JSEG-Deng and Manjunath [99]	Consisting of two independent steps: color quantization and spatial segmentation. In the first step, image colors are quantized in different classes, which are used to create an image class map. The image segmentation results from the application of a region growing method to the set of multiscale images, formed through the application of the class map based segmentation evaluation criterion.
CTM-Yang et al. [100]	Colour–texture features at pixel level are extracted simultaneously by stacking the intensity values within a 7x7 window for each band of the CIE Lab converted image. Segmentation is formulated as a data clustering process. To reduce the dimensionality of the colour–texture vectors, Principal Component Analysis is used. To overcome the difficulty related to the fact that often the colour–texture information cannot be described with normal distributions, a coding-based clustering algorithm is employed that is able to accommodate input data defined by degenerate Gaussian mixtures.
Chen et al. [101]	Segmentation of natural images into perceptually distinct regions with application to content-based image retrieval. Local colour features are extracted using a spatially Adaptive Clustering Algorithm. Texture features are computed through a multi-scale frequency decomposition procedure. Colour and texture features are integrated using a region growing algorithm that generates a primary segmentation that is improved through a post-processing step that implements a border refinement procedure.
Han et al. [102]	A segmentation framework developed to identify the foreground object in natural colour images. Colour features are extracted from the CIE Lab converted colour image. Texture features are computed from the luminance component of the input image using the multi-scale nonlinear structure tensor. To reduce the dimensionality of the colour–texture feature space, the colour information is clustered using a binary tree quantisation procedure and the features in the texture domain are clustered using a K-Means algorithm. The resulting colour and texture features are modelled by Gaussian Mixture Models and integrated into a framework based on the GrabCut algorithm. The accuracy of the algorithm is improved by an adaptive feature integration strategy that consists of adjusting a weighting factor for colour and texture in the segmentation process.
GrabCut-Rother et al. [103]	A graph-cut approach extension, with a simpler user interaction and an iterative version of the optimization method. An algorithm for “border matting” is used to estimate simultaneously the alpha-matte around an object boundary and the colours of foreground pixels.
Blobworld-Carson et al. [104]	The goal of the proposal is to partition the input image in perceptual coherent regions. It includes an isotropy, polarity and contrast features in a multi-scale texture model. Colour features are extracted on an independent channel from the CIE Lab converted image previously filtered with a Gaussian operator. For automatic colour–texture image segmentation, the distribution of the colour, texture and position features are jointly modeled using Gaussian Mixture Models. The Blobworld algorithm is able to segment the image into compact regions, being suitable to integrate a content-based image retrieval system.
CTex-Ilea and Whelan [105]	Colour and texture are treated on separate channels. Colour segmentation involves the statistical analysis of data using multi-space colour representations. After filtering the input data using a Gradient-Boosted Forward and Backward anisotropic diffusion algorithm, the colour segmentation algorithm extracts the dominant colours and identifies the optimal number of clusters using an unsupervised procedure based on a Self Organising Map network. After, the image is analysed in a complementary colour space where the number of clusters previously calculated performs the synchronisation between the two computational streams of the algorithm. Finally, clustered results obtained for each colour space form the input for a multi-space clustering process that outputs the final colour segmented image. The extraction of the texture features from the luminance component of the original image uses a multi-channel texture decomposition technique based on Gabor filters. The colour and texture features are integrated in an Adaptive Spatial K-Means framework that partitions the data mapped into the colour-texture space by adaptively sampling the local texture continuity and the local colour smoothness in the image.
Malik et al. [69]	An algorithm for partitioning grayscale images into disjoint regions of coherent brightness and texture, where cues of colors and texture differences of natural images are exploited simultaneously. Contours are treated in the intervening contour framework, while texture is analysed using textons. Given the different domain of applicability of each cue, a gating operator is introduced based on the texturedness of the neighbourhood at a pixel. Given a local measure of the similarity between nearby pixels, the spectral graph theoretic framework of normalized cuts is used to find partitions of the image in regions of coherent texture and brightness.