Induction operators for a computational colour–texture representation
Introduction
Any scene of the world is projected on our retina as a map of different regions that are the projections of 3D surfaces. The properties of these projections are derived from the position and orientation of the surfaces in the scene, the observer location, and the light that provoke the neuronal excitation of the visual system. In computer vision, people usually deal with a set of surface properties, shape, orientation, colour, and texture. In this work we will only deal with the last two and their dependency on the final perception.
Although both are inherent properties of surfaces, these two visual cues have usually been studied separately [3], [4]. This is probably due to their usual representations, while colour is a point feature given by the value of a pixel in several bands or channels, texture has to be modelled as a spatial relationship of the point with its neighbours. In Fig. 1 we see the RGB channels of a colour image, we can observe that the spatial information perceived from the colour image does not appear as it is in any of the channels. Therefore specific representations have to be built in order to deal with both cues at the same time.
The study of colour–texture representations has received increasing attention. The objective of many researchers has been to find co-joint representations of spatial and chromatic information which capture the spatial dependencies (in particular, correlation) within and among spectral bands. One of the most frequent approaches is to define a feature vector joining grey level texture features and colour features [5], [6]. Another one is to extend classical texture models, such as Markov Random fields and the autocorrelation function, in order to deal with multichannel images [7], [8], or wavelet analysis extended to colour images by combining the results in colour channels [9]. Other works, like [10], convert RGB values into a single code from which texture measurements are computed as if it were a grey scale image. Spatio-chromatic representations are computed in [11], [12] over the smoothed Laplacian of the image, and the structural tensor that is usually used to represent local texture properties is extended to colour images in [13].
However, we want to highlight the approach presented in [2], [14] that is based on known perceptual mechanisms of the human visual system. Colour–texture interaction is represented as a perceptual blurring which depends on the spatial frequency of the coloured patterns and the observer position. This approach is based on important conclusions from psychophysical works on colour–texture interaction [1], [15], [16], [17], [18]. We will review the details of this perceptual blurring in Section 5.
Following on from this prior work, in this paper we propose a global perceptual approach of colour–texture representation that combines the introduced perceptual blurring with a first approach to a perceptual sharpening. The combination of these two operators can simulate the visual process of colour–texture perception with the colour induction mechanisms produced by different spatial frequencies.
To this end, in 2 Colour, 3 Texture we give, separately, the basis of computational representations of colour–texture. In Section 4 we briefly introduce the colour induction mechanisms and their relationship to spatial frequency. Section 5 explains the essentials of perceptual blurring that simulates chromatic assimilation, and in Section 6 we introduce a perceptual sharpening that simulates the colour contrast phenomenon. In Section 7 we consider colour–texture perception as a visual process that integrates both induction mechanisms, this will be the basis for a global representation of colour–texture that is introduced in Section 8. Finally, in Section 9 we discuss some properties of the inductors operators and the advantages of using them in computing textural properties and in Section 10 we summarise the open problems that still remain in order to complete the proposed computational model.
Section snippets
Colour
Colour is the visual cue derived from the human visual processing of the electromagnetic radiation that reaches the retina. This process can be seen as a change in representation, which, in general, implies a dimensionality reduction. Although colour was not given much importance in the first decades of computer vision, since most previous work has been undertaken on grey-level images, the situation has changed and colour has become a very important visual cue for most of the vision tasks, such
Texture
Texture is the visual cue derived from non-homogeneous surfaces in scenes. Depending on the surface reflectance, positioning of the observer, and lighting conditions, we can obtain different texture images from the same surface. Although there are some recent works dealing with the recovery of the physical reflectance properties of a texture [26], [27] and some other works that have recovered 3D shape information from texture [28], [29], the most traditional approach in computer vision has been
Colour induction
Colour induction are the colour phenomena that changes the colour appearance of a stimulus due to the influence of the scene contents in the field of view. There are different types of induction phenomena such as colour adaptation, colour assimilation or colour contrast amongst others.
Colour adaptation is involved in any scene interpretation and occurs when scene colours are perceived without being affected by the illuminant influence, this ability presented by the HVS has been modelled in
Colour assimilation as a perceptual blurring
A computational model of colour assimilation has already been given by an isotropic blurring of an image on the opponent-colour space, this has been proposed by Petrou et al. [2] as a perceptual blurring. Colour assimilation effects were first measured by Wandell et al. [1], where they take perceptual measurements of quality on printed patterns. To achieve this goal, they propose the Spatial-CIELAB space that is given by a two-steps process: an opponent colour channel transformation and a
Colour contrast as a perceptual sharpening
As we have already explained in Section 4, colour contrast is a complementary mechanism to colour assimilation. Colour contrast arises on regions of low spatial frequency and shifts the chromaticity of the stimulus in a direction away from the chromaticity of the surround. In this section we define a computational operator that simulates the colour contrast phenomenon.
This operator enhances differences in the transitions among colours of regions presenting low frequencies. While the
Colour–texture perception as a visual process
In 2 Colour, 3 Texture we have explained, respectively, how colour and texture can be computationally represented as separate cues. However, we need to work on images from real world, which is neither a grey world, nor a Mondrian world.
In Section 4 we have briefly introduced the colour induction phenomenon that explains how colour changes depend on spatial frequencies, and how opponency can represent these interactions. These perceptual considerations can be simulated with two types of
A computational colour–texture model
Considering the assumptions done in the previous section, we now define the complete perceptual tower which we intend as a general perceptual representation of coloured textures. Therefore, for a given image we can build a perceptual tower representing colour information as it was observed from different observer distances. It can be considered as a colour–texture front-end representing colour–texture interaction and the basic step for further processing, a parallel approach to a scale-space
Discussion
In this paper we outlined a framework to deal with the interaction between texture and colour in a perceptual sense. Our contribution extends previous working [2] by incorporating into the perceptual tower an inductor operator, perceptual sharpening, which models the effect of chromatic contrast (a complementary effect to assimilation which is modelled by perceptual blurring).
Combining these two operators we can build a global framework that acts as a primal sketch for colour representation.
Future work
The first step to provide a full perceptual approach will require a psychophysical measurement of chromatic contrast, this will imply a tedious task of psychophysical experimentation. Once it is done, we will have to match the measurements with the sharpening parameters.
In the work presented we have taken all the attention on getting induction operators presenting good properties to represent colour appearance, however we have not taken attention to the computational cost of these operators. In
Acknowledgements
This work has been funded by project TIC2000-0382 of the MCYT of the Spanish Goverment.
References (60)
- et al.
A review of recent texture segmentation and feature extraction techniques
CVGIP: Image Understanding
(1993) - et al.
On the classification of image regions by colour, texture and shape
Pattern Recognition
(1993) Coherence-enhancing diffusion of colour images
Image Vision Comput.
(1999)- et al.
Color based object recognition
Pattern Recognition
(1999) Recovering surface shape and orientation from texture
Artificial Intell.
(1981)Shape from texture and contour by weak isotropy
Artificial Intell.
(1993)Texture analysis—a survey
Signal Process.
(1980)- et al.
Identifying high level features of texture perception
CVGIP: Graphical Models Image Process.
(1993) - et al.
Spatial frequency channels and perceptual grouping in texture segregation
Computer Vision Graphics Image Process.
(1987) - et al.
Chromatic induction: border contrast or adaptation to surrounding light?
Vision Res.
(1998)
Chromatic induction with remote chromatic contrast varied in magnitude, spatial frequency, and chromaticity
Vision Res.
A central mechanism of chromatic contrast
Vision Res.
The role of spatial frequency in color induction
Vision Res.
The consequences of opponent rectification: the effect of surround size and luminance on color appearance
Vision Res.
Color appearance: neutral surrounds and spatial contrast
Vision Res.
Pattern–color separable pathways predict sensitivity to simple colored patterns
Vision Res.
Photoreceptor sensitivity changes explain color appearance shifts induced by large uniform backgrounds in dichoptic matching
Vision Res.
On the computation of a scale-space primal sketch
J. Visual Commun. Image Representation
Perceptual smoothing and segmentation of colour textures
The synthesis and analysis of color images
IEEE Trans. Pattern Anal. Machine Intell.
Markov random field models for unsupervised segmentation of textured color images
IEEE Trans. Pattern Anal. Mach. Intell.
Illumination-invariant recognition of texture in color images
J. Opt. Soc. Am.
Color constant color indexing
IEEE Trans. Pattern Anal. Mach. Intell.
Segmentation of color textures
IEEE Trans. Pattern Anal. Mach. Intell.
Color appearance of mixture gratings
Vision Res.
Cited by (11)
Influence of normalization and color space to color texture classification
2017, Pattern RecognitionCitation Excerpt :Statistical and spectral techniques are the most popular in the literature of texture classification. Structural approach usually considers the texture as a composition of texture primitives [19] and it only performs well on very regular textures. Model-based approaches, such as the use of Markov Random Fields (MRF) [20], are not widely extended for image classification [21].
Application-independent feature selection for texture classification
2010, Pattern RecognitionCitation Excerpt :In fact, the work presented by Rao and Lohse in [17] is considered as a pioneering reference in the literature related to perceptual texture analysis, since they established the basis for a standard computational representation of texture. In this way, that study has been extensively referenced in the literature by recent works on texture description and representation (e.g. [19,20]), texture retrieval (e.g. [21]), texture synthesis (e.g. [22]), and color texture segmentation (e.g. [23]). Feature selection for image classification based on perceptual criteria is an incipient research field that has started to receive considerable attention in the literature during the last recent years, mainly for the classification of whole images or individual pixels, or to search images in databases by content, although not necessarily using texture as the only visual cue (e.g. [24,25]).
Parametric fuzzy sets for automatic color naming
2008, Journal of the Optical Society of America A: Optics and Image Science, and Vision