Lattice estimation from images of patterns that exhibit translational symmetry☆,☆☆
Introduction
Regular textures can be modelled as consisting of repeated texture elements, or texels. The texels tesselate (tile) the image or, more generally, a surface. Varying illumination, varying physical characteristics of the textured surface, geometric deformations and sensor noise all result in images of such patterns exhibiting approximately regular, as opposed to exactly regular, texture. In this paper, we consider the task of automatically extracting texels from images. In particular we are interested in analysing patterns in printed textiles. Such patterns often exhibit translational symmetry and the visual structure within each texel can itself be complex. Translationally symmetric regular textures can always be generated by a pair of shortest vectors (two linearly independent directions), t1 and t2, that define the size, shape and orientation (but not the position) of the texel and the lattice that it generates. The lattice topology is always then quadrilateral. Such textures form one class of wallpaper pattern [1]. We restrict ourselves to consideration of images of planar, approximately regular textures viewed under orthographic projection. Whilst this might at first seem restrictive, this problem is, as will become apparent, far from completely solved. Furthermore, solutions will find application in analysis, retrieval and restoration of textile, wallpaper and tile design images, for example. The aim then, is to automatically decide whether a translationally symmetric regular texture class provides a good model for a texture image, and in cases in which it does, to extract the most predictive texel geometry.
In an earlier conference paper [2], this problem was cast in a statistical model comparison framework; models representing different hypotheses for lattice geometry were compared using the Bayesian Information Criterion (BIC). In this paper we extend this work in several ways. A more extensive discussion of related literature is provided setting the work more clearly in context, and certain aspects of the method have been clarified. False hypotheses are pruned using a constraint on the angle between t1 and t2. Classification of texture images as regular or irregular using model comparison is compared empirically to an alternative method [3], [4]. Model comparison is performed using the Akaike Information Criterion (AIC) as well as BIC. New experiments are reported on larger datasets, and results are reported separately for private and public datasets thus enabling future comparisons to be made using the public data. A detailed comparison of lattices obtained using several methods is presented. Experiments involving qualitative observer evaluations are expanded. The proposed method and implementations of two alternative methods from the literature are compared quantitatively against ground-truth annotations using two quantitative evaluation measures.
The rest of this paper is organized as follows. Section 2 outlines the most closely related previous work. Section 3 presents the model comparison framework. Section 4 describes details of lattice models. Section 5 describes the method used in our experiments for generating lattice hypotheses. Evaluations are described in Section 6 and conclusions are drawn in Section 7.
Section snippets
Related work
Previous work proposed for texel and lattice extraction can be grouped broadly into two categories: the local feature-based approach [5], [6], [7], [8], [9], [10], [11], [12], [13] and the global structure-based approach [1], [14], [15], [16], [17], [18]. All texture analysis is necessarily both local and global. The categorisation is in terms of the computational approach: whether it starts by identifying local features and proceeds to analyse global structure, or it starts with a global
Model comparison framework
The approach we take is to formulate texel hypotheses as statistical models and then compare these models given the image data. It is not sufficient for a model to be able to fit the data well. The best texel hypothesis under this criterion would be the image itself whereas our purpose is to extract the smallest texture element. Therefore, overfitting must be guarded against by penalising model complexity. Texel hypothesis comparison can be regarded as a model comparison problem for
Lattice models
Each of the Q clusters was modelled as Gaussian, a modelling choice motivated by exploratory data analysis which showed Gaussian-like distributions, as well as by parsimony and computational simplicity. Experimental results (see Section 6) suggest that this was a reasonable choice, at least for the data used here. However, when datasets contain many spatially localised disruptions (due to physical surface damage, local photometric variations such as specular highlights, local geometric
Lattice hypothesis generation
In principle there is an infinite number of lattice hypotheses. However, probability density will be highly concentrated at multiple peaks in the hypothesis space. The posterior distribution can therefore be well represented by only considering a, typically small, number of hypotheses at these peaks. In the maximum a posteriori setting adopted here, the approach taken is to identify multiple hypotheses in a data-driven manner and then compare these hypotheses using BIC. The approach is general
Evaluation
There are a few systematic evaluations of lattice extraction algorithms in the literature [29], [30], [19]. Chen et al. [29] reported an evaluation of translational symmetry in terms of the number of texels but not in terms of their shape. Park et al. [19] described an evaluation method for detection of deformed lattice regions in photographs of real-world scenes. Ground-truth lattice structures were annotated using an interactive editing tool. Detection rate was defined as the number of valid
Discussion and conclusions
It seems clear that the method proposed in this paper has superior accuracy to the methods with which it was compared. This was the case on both the commercial and public databases. In particular, the number of lattice extractions not judged to be clearly correct by observers was only a third of that obtained by the closest competing method. Even with parameter settings which disadvantage it relative to the other methods, the proposed method gave markedly better lattice estimates (Table 2).
Acknowledgements
The authors are grateful to Ruixuan Wang and Annette Ward for their helpful discussions. They would like to thank J. Hays and M. Park for providing their source code. This research was supported by the UK Technology Strategy Board grant “FABRIC: Fashion and Apparel Browsing for Inspirational Content” in collaboration with Liberty Art Fabrics, System Simulation Ltd., and the Victoria & Albert Museum.
References (33)
- et al.
An efficient computation method for the texture browsing descriptor of MPEG-7
Image Vis. Comput.
(2005) - et al.
Extracting periodicity of a regular texture based on autocorrelation functions
Pattern Recogn. Lett.
(1997) Pattern regularity as a visual key
Image Vis. Comput.
(2000)On indexing the periodicity of image textures
Image Vis. Comput.
(2001)- et al.
An evolutionary system for near-regular texture synthesis
Pattern Recogn.
(2007) - et al.
A computational model for periodic pattern perception based on frieze and wallpaper groups
IEEE Trans. Pattern Anal. Mach. Intell.
(2004) - et al.
Regular texture analysis as statistical model selection
- et al.
A new method for coarse classification of textures and class weight estimation for texture retrieval
Pattern Recogn. Image Anal.
(2002) - et al.
Discovering texture regularity as a higher-order correspondence problem
- et al.
A lattice-based MRF model for dynamic near-regular texture tracking
IEEE Trans. Pattern Anal. Mach. Intell.
(2007)
Detecting, localizing and grouping repeated scene elements from an image
Noncombinational detection of regular repetitions under perspective skew
IEEE Trans. Pattern Anal. Mach. Intell.
Geometric grouping of repeated elements within images
Shape from texture without boundaries
Recovering shape and irradiance maps from rich dense texton fields
Deformed lattice discovery via efficient mean-shift belief propagation
Cited by (8)
Fabric defect inspection based on lattice segmentation and template statistics
2020, Information SciencesCitation Excerpt :The former gets the information about the texture primitives before inferring the placement rule, while the latter infers the rule first and then obtains the texture primitives based on the rule. Generally, global feature-based methods are suitable for images containing large texture primitives of sophisticated appearances [13]. Consequently, global feature-based methods are commonly adopted in fabric inspection, e.g., lattice segmentation method in [16–18,23].
Fabric defect inspection based on lattice segmentation and lattice templates
2018, Journal of the Franklin InstituteCitation Excerpt :The former identifies the texture primitives before inferring the placement rule, while the latter reverses the order. Generally, GFM are suitable for images containing large texture primitives of sophisticated appearances [27]. Consequently, GFM are commonly adopted in AFI, e.g., Liu’s et al method in [28] and LS method in [17] and [18].
Fabric defect inspection based on isotropic lattice segmentation
2017, Journal of the Franklin InstituteFabric defect inspection based on lattice segmentation and Gabor filtering
2017, NeurocomputingCitation Excerpt :The following literature review shortly discusses each component related with fabric defect inspection. The lattice segmentation may be categorized based on their computational approaches [25], i.e., the local feature-based and the global structure-based approaches. The former identifies the local features first and then analyzes the global structure, the latter conducts the two steps in a reverse order.
Fabric defect fetection via weighted low-rank decomposition and Laplacian regularization
2020, Journal of Engineered Fibers and FabricsAnalysis of Planar Ornament Patterns via Motif Asymmetry Assumption and Local Connections
2019, Journal of Mathematical Imaging and Vision
- ☆
This paper has been recommended for acceptance by Enrique Dunn, Ph.D.
- ☆☆
This research was carried out whilst J. Han was with the School of Computing, University of Dundee, DD1 4HN, UK.