Segmentation of images is one of the mandated and crucial steps in analyzing the images while accomplishing the machine vision tasks for targeted performance. It has multiple applications in numerous fields such as image data retrieval, pattern analysis, industrial and medical image analysis, machine learning and various other areas. Segmentation of images is performed using various techniques and approaches such as thresholding, edge-based methods, clustering-based methods, level set method, parametric methods etc. Segmentation, based on Image Texture as a cue, is a perpetually popular research domain. Due to many texture attributes, researchers have not been able to converge upon a universal definition of texture, thus making the texture segmentation task as non-trivial. Haralick  define texture as repetition of grey levels in a neighbourhood and have categorized the texture as regular, irregular, coarse or fine. A strong correlation between human vision and texture has motivated various research activities in the domains of synthesis of textures, classification of textures and shape extraction using textures  leading to the applications such as 3D imaging, synthesis of natural scenes and computer graphics, extraction of surface shape from textures. Some of the texture analysis applications are diagnoses using X-ray images, retrieval of images from databases, weather forecasting, sea-ice imagery analysis etc.
An extensive literature review on texture segmentation, for the published work over the period of more than past three decades, segmentation of textures using Gabor filters with traditional classifiers [8, 9, 13, 25, 31, 41, 42] has been discussed, in majority. Segmentation of textured SAR and medical images using Markov Random Fields (MRF) have been proposed in [28, 29, 40, 41]. The color and texture-based image analysis is applied recently by researchers for medical image analysis [5, 6, 19, 39] industrial applications [17, 18, 34], remote sensing [49, 51], analysis of animal audio signals  and saliency estimation in the set of video frames and images . All these researchers carried out experimentation on non-degraded noise-free texture images [23, 29]. However, noise is the important attribute that affects the performance of segmenting algorithms and also an integral part of measurements, thus making it imperative to test the noise robustness of segmenters for textures corrupted with additive Gaussian, Poisson, salt pepper and speckle noise, in several signal-to-noise-ratio (SNR) steps . Identifying the due importance of noise in segmenting algorithms, as per the most recent literature [1, 12, 47], researchers have initiated the work on noisy textures to address noisy texture-based image retrieval and noisy texture analysis. They have used four texture datasets, namely, Outex, Broadatz, STex and Curet for experimentation. Though there are 30 plus texture datasets available globally as per [23, 33] for researchers to experiment with, the experimentation of noisy textures is not yet explored to a deeper extent by the researchers encompassing these datasets.
As per recent literature  Prague texture benchmark offers potentially unlimited texture data for experimentation and it is developed to evaluate objective performance and ranking of texture segmentation algorithms and it contains largest variety of multi-class textures, that makes it the default choice of the researchers across the globe for relevant experimentation [29, 33]. As per state-of-the-art, more than 11 contemporary approaches of texture segmentation have been applied on the Prague texture dataset , though none of these algorithms used benchmark images degraded with noise. Many of the researchers have used Brodatz textures for experimentation as per the literature review, therefore proposed approach is also evaluated on color Brodatz texture benchmark images in . In many of the real-life scenarios, salt pepper and Gaussian noise are the significant sources of corruption for texture images caused during acquisition, coding or transmission [3, 20, 48]. The researchers have also reported their work related to applications such as texture-based automatic quality evaluation of vegetables and fruits , diagnosis of thyroid cancer images  and recognition of facial expressions from the idea of the human face , after restoring the noisy images. As per literature review, most of the researchers focused on improvement in texture segmentation accuracy and real-world problem of noisy texture segmentation is not addressed to deeper extent. Our work on investigating the segmentation of noisy color textures from the Prague and Brodatz texture benchmark images, is thus duly motivated from such latest reported works of peer researchers round the globe. Proposed approach ensured unparallel performance with respect to rotational and scale variance although that has not been targeted as the immediate research concern.
In this work, texture images are degraded with salt pepper and Gaussian noise, respectively. Texture images degraded with Gaussian noise are restored using a colour block matching 3D (C-BM3D) algorithm due to its best restoration performance among the latest 14 algorithms [10, 11, 15, 30, 35], whereas, de-noising of texture images corrupted with salt pepper noise is achieved using a nature and bio-inspired algorithm, namely, Cellular Automata (CA) [14, 26] that being the best performing among 15 recent algorithms for this noise [14, 15, 35] and further it eliminates limitations of the Median filter and its variants, while preserving the edge information [14, 24, 32, 44].
A two-stage novel MRF segmentation algorithm using the custom Median filter is developed and used for segmentation of Prague texture images and Brodatz texture images. C-BM3D and CA restoration algorithms with the novel MRF based segmentation approach give better segmentation performance than deep learning-based approach using U-net in  on Brodatz texture benchmark images and recent texture segmentation algorithms in  and deep learning approach using fully convolutional network (FCNT) reported in  on Prague texture images, to compare with a few.
We could attain higher mean segmentation accuracy on tainted texture images than recent approaches evaluated on noise free texture images. This approach is evaluated on two color Brodatz texture benchmark images in , and accuracy achieved is higher on tainted Brodatz texture images than deep learning approach in , evaluated on noise free Brodatz texture benchmark images. Our segmentation accuracy achieved is higher by 16.04% for salt pepper noise with 70% noise density and it is higher by 15.10% than recent approaches for Gaussian noise with variance of 50 on 16-class Brodatz texture benchmark image. Our segmentation accuracy achieved is higher than the recent approach proposed by Kiechle et al.  for Prague benchmark images. It is higher than 8% for Gaussian noise with variance of 50 and it is higher by 4.47% for salt pepper noise with 70% noise density. Also, we could restore texture benchmark images in the Prague dataset and Brodatz texture benchmark images, degraded with salt pepper noise and Gaussian noise, with best performing algorithms viz. CA and C-BM3D, respectively, with an achievement of excellent mean values of performance metrics, structural similarity index (SSIM) and peak signal to noise ratio (PSNR). The proposed color texture segmentation approach can be applied for satellite, noisy texture-based retrieval and analysis, for river scene segmentation , diagnoses of breast  and prostate  cancer, analysis of radiographic images in dentistry , analysis of human hair and skin images useful for pharmaceutical and cosmetic industries , classification of fabrics, rocks and for inspection of aircraft surfaces, fruits, steel and wood .
The paper is further organized as follows: Restoration algorithms are discussed in Section-2. Segmentation problem formulation and segmentation approach are detailed in section-3. Experimentation and the obtained results are discussed in section-4. Section-5 reports the conclusion.