Local directional ZigZag pattern: A rotation invariant descriptor for texture classification
Introduction
Texture is ubiquitous in natural images that carries fundamental characteristic of appearance of all natural surfaces. Texture classification is one of the active and challenging problems in texture analysis. It has drawn a lot of attention during the past decades as it plays a crucial role in the area of pattern recognition and computer vision. It has a wide range of applications such as medical image analysis, remote sensing, fabric inspection, segmentation and content-based image retrieval [1], it also includes iris based biometric recognition. Since feature extraction is often performed locally based on regions of the neighborhood, most research efforts have been directed to various local neighborhood patterns of an image. An important issue is how to represent the texture effectively. Basically, texture representation can be categorized in terms of the employed approaches, i.e. geometrical, structural, model-based, statistical, and signal processing. Earlier texture classification methods focus on the statistical analysis of texture images which include the co-occurrence matrix based approach [2] and filtering based techniques [3]. These methods provide good classification performance as long as both training and test sample images have identical orientations. However, arbitrary rotations which could occur in a real-world scene, affect the performance of statistical methods. Therefore, rotation invariance is a crucial issue to be addressed and attention has been focused on the design of geometrically and photometrically invariant local texture representation [4], [5], [6], [7], [8], [9]. This paper concentrated on the problem of rotation invariant texture representation.
The extraction of rotation invariant features is usually a complex process where spacial care is required in intermediate step which is computationally demanding [10]. The literature of computer vision shows the work on this aspect has started in the nineties of last century [11]. Kashyap and Khotanzad first proposed circular autoregressive dense approach [12] for the rotation invariance texture classification. Many models have been explored for rotation invariance texture classification, including multi-resolution [13], hidden Markov model [14], and Gaussian Markov model [15]. Recently Varma and Zisserman proposed VZ-MR8 [5] to learn a texton dictionary from a set of training images which are rotation and scale invariant and then classified the unknown sample images using learned texton distributions. Later, Varma and Zisserman proposed VZ-Patch [7], another texton based algorithm to represent the feature distribution, which extracts texton from local image intensity directly. The downside of these methods are feature extraction and matching times which is not very favourable.
In 1996 a simple and computationally efficient texture representation, called local binary pattern (Lbp) was proposed by Ojala et al. [16], which is invariant to the uniform intensity changes. The Lbp has been widely used in other domains such as texture segmentation, face recognition, shape localization and object recognition [17]. Variants of Lbp have been proposed due to immense success of Lbp in pattern recognition and computer vision problems. Liaor et al. [18] proposed Dlbp, a dominant pattern by encoding only the most frequently occurred patterns (around 80%) to improve the performance. Guo et al. introduced Lbp variance (Lbpv) [19] and completed Lbp (Clbp) [20] to enhance the descriptive power and improve the texture classification, for effective recognition of face images, Zhang et al. [21] used local derivative pattern (Ldp) with higher order and showed that Ldp performs much better than conventional Lbp. Dubey et al. introduced Lwp [22] to find the histogram feature vector for biomedical image retrieval. Guo et al. introduced a local directional derivative pattern (Lddp) [23] which includes the directional information with Ldp for rotation invariant texture classification. Tan and Triggs introduced local ternary pattern (Ltp) [24] and Liu et al. proposed [25] noise tolerant descriptor to improve the texture classification performance under varying illumination and noisy conditions. Mehta and Egazarian proposed a variant of Lbp so-called Drlbp descriptor for rotation invariant texture classification. Recently, Roy et al. [26] introduced a complete dual-cross pattern (Cdcp) to address the scale and rotational effects in unconstrained texture classification.
However, most of the descriptors are based on the same basic idea of Lbp and extracts only circular isotropic micro structure of the texture image which is not enough to describe the texture information and do not sufficiently address the rotation invariant issues. Hence, inspired by the ZigZag scanning of discrete cosine transform (Dct) [27] coding technique, a new image descriptor called Local Directional ZigZag Pattern (Ldzp) is proposed for effective texture representation and classification. The readers should not confuse the ZigZag scanning of Dct encoding which perform in frequency domain for data compression, with our ZigZag scanning of Ldzp, performed in spatial domain to select the ordering for generating the descriptor. The schematic diagram of proposed Ldzp based texture classification framework is shown in Fig. 1.
The main contributions of this letter are as follows: We propose a Local ZigZag Pattern (Lzp) where sampling points fall exactly at the integer pixel position and avoids inaccuracy of interpolation of gray values, characterizes the spatial local ZigZag structure of a texture image; The Lzp presents a strong angular relation between two consecutive pixels with respect to the center as well as two alternate pixels with respect to their intermediate pixel which makes it more efficient to capture non-uniform local texture pattern; To compute the Ldzp descriptor we calculate local directional edge map (Ldem) of a texture image using the Kirsch compass mask in six different directions and encode all directional edge responses using Lzp which reduces the noise effects; The uniform pattern histograms are computed from all directional Lzp maps to reduce the dimension. Finally all the histograms are concatenated to create the Ldzp descriptor, which encodes the directional Lzp pattern information and makes the descriptor invariance to rotation.
The rest of the letter is organized as follows. Section 2 describes the details of proposed local ZigZag pattern (Lzp). The proposed local directional ZigZag pattern (Ldzp) descriptor is presented in Section 3. The effectiveness of the Lzp over Lbp are discussed in Section 4. Section 5 describes the evaluation criteria. The experimental results are summarized in Section 6 and the conclusion is drawn in Section 7.
Section snippets
Local ZigZag pattern
This work introduces a novel and efficient texture descriptor from the relation between a center pixel and its local neighboring pixels by ZigZag scanning, called local ZigZag pattern (Lzp). Hence Lzp is a local gray scale texture descriptor which represents the local spatial ZigZag structure of a texture image as shown in Fig. 2(a). Let in a gray-scale image I, be the center pixel of a 3 × 3 local neighboring window having gray value and the nth neighbors of are denoted
Local directional ZigZag pattern
The Lbp represents a non-directional first order circular derivative of local texture pattern which labels each pixel by thresholding a set of sampled point of its even space circular neighbourhood. It encodes the local micro-information as a binary string without considering the suitable neighboring relationship. Whereas Ldp extracts the more detailed local textural information with the higher order directional derivative variation of each pixel neighborhood. However, this method marks only
Local ZigZag pattern vs. local binary pattern
In conventional Lbp, the neighboring sampling points that do not fall exactly within the integer pixel positions have been estimated by bi-linear interpolation or rounding operation which leads to the unreliable texture information due to inaccuracy of interpolation or rounding operation. In case of Lbp the feature dimension exponentially increases with the number of sample points and it leads to difficulties in both computation and classification performance. However, the proposed Lzp replaces
Similarity matching using Ldzp
In this work, the texture classification is performed via non-parametric Nnc classifier. The Nnc with χ2-distance [4], [7], [20] is used to show the effectiveness of the proposed Ldzp descriptor. Two histograms and are compared using χ2 distance, defined as follows, where M represents the total number of bins, H1 and H2 represent the extracted features of a trained model and test sample. The class of test sample H1 is assigned to the class of
Results and discussion
To figure out the texture classification performance of the proposed Ldzp descriptor, experiments are carried out on two large and commonly used well-known Outex_TC_00010 (TC10) and Outex_TC_00012 (TC12) [28] texture databases. These databases contain 24 classes of homogeneous texture images of size 128 × 128 pixels. Outex_TC_00010 (Outex_TC10) contains texture images under illuminant “inca” whereas Outex_TC_00012 (Outex_TC12) contain texture images with 3 different illuminants (“inca”,
Conclusion
In this letter, we proposed a novel and efficient descriptor for rotation invariant texture image classification by exploring the local directional ZigZag pattern (Ldzp). To compute Ldzp descriptor, at first directional edge response of a texture image is obtained using Kirsch compass mask in six different directions. Then the proposed local ZigZag pattern (Lzp) which characterize local spatial ZigZag structure of texture is used to encode directional edge responses. Finally, Ldzp feature
Acknowledgements
The authors would like to thank Machine Vision Group, University of Oulu, Finland and Visual Geometry Group, University of Oxford, UK for sharing the program codes of Lbp and VZ_MR8.
References (40)
- et al.
Brief review of invariant texture analysis methods
Pattern Recognit.
(2002) - et al.
Texture classification and segmentation using multiresolution simultaneous autoregressive models
Pattern Recognit.
(1992) - et al.
A comparative study of texture measures with classification based on featured distributions
Pattern Recognit
(1996) - et al.
Rotation invariant texture classification using lbp variance (lbpv) with global matching
Pattern Recognit.
(2010) - et al.
Noise tolerant local binary pattern operator for efficient texture analysis
Pattern Recognit. Lett.
(2012) - et al.
Dominant rotated local binary patterns (drlbp) for texture classification
Pattern Recognit. Lett.
(2016) An introduction to roc analysis
Pattern Recognit. Lett.
(2006)- et al.
Face recognition using scale-adaptive directional and textural features
Pattern Recognit.
(2014) - et al.
Texture analysis
Handbook Pattern Recognit. Comput. Vis.
(1993) - et al.
Textural features for image classification
IEEE Trans. Syst. Man Cybern.
(1973)
Filtering for texture classification: a comparative study
IEEE Trans. Pattern Anal. Mach. Intell.
Multiresolution gray-scale and rotation invariant texture classification with local binary patterns
IEEE Trans. Pattern Anal. Mach. Intell.
A statistical approach to texture classification from single images
Int. J. Comput. Vis.
Local features and kernels for classification of texture and object categories: a comprehensive study
Int. J. Comput. Vis.
A statistical approach to material classification using image patch exemplars
IEEE Trans. Pattern Anal. Mach. Intell.
Using basic image features for texture classification
Int. J. Comput. Vis.
Texture classification from random features
IEEE Trans. Pattern Anal. Mach. Intell.
Geometric invariants and object recognition
Int. J. Comput. Vis.
A model-based method for rotation invariant texture classification
IEEE Trans. Pattern Anal. Mach. Intell.
Rotation and gray scale transform invariant texture identification using wavelet decomposition and hidden Markov model
IEEE Trans. Pattern Anal. Mach. Intell.
Cited by (43)
Texture and material classification with multi-scale ternary and septenary patterns
2023, Journal of King Saud University - Computer and Information SciencesCervical cancer diagnosis based on modified uniform local ternary patterns and feed forward multilayer network optimized by genetic algorithm
2022, Computers in Biology and MedicineCitation Excerpt :The LBPP,R operator produces (2P) different output values, corresponding to the 2P different binary patterns that can be formed by the P pixels in the neighbor set and radius R. Various ideas have been presented to improve the performance of LBP, which are known as lbp-like descriptors. In some cases, such as one dimensional local binary patterns (1DLBP) [17] and local zigzag patterns [18], neighborhood selection has changed. In some cases, such as local ternary patterns (LTP) [19], median binary patterns (MBP), innovative ideas have been proposed to determine binary patterns.
Grayscale-inversion and rotation invariant image description using local ternary derivative pattern with dominant structure encoding
2022, Expert Systems with ApplicationsCitation Excerpt :Since the images in these four databases have no obvious grayscale-inversion changes, the test images are pre-processed to simulate such changes. We compare state-of-the-art LBP and LTP based descriptors including LBP (Ojala, Pietikäinen & Mäenpää, 2002), LTP (Tan & Triggs, 2010), CLBP (Guo et al., 2010), SILTP (Liao et al., 2010), LDTP (Ramírez Rivera et al., 2015), DRLBP (Satpathy et al., 2014), DRLTP (Satpathy et al., 2014), MRELBP (Liu et al., 2016), LOOP (Chakraborti et al., 2018), LGP (Jun et al., 2013), GLBP (He & Sang, 2011), NRLBP (Nguyen et al., 2010), SLGP (Song, Xin, Gao, Zhang & Zhang, 2018), CGRI-LBP (Song, Xin, Luo & Gao, 2018), and LDZP (Roy et al., 2018). The CUReT database5 contains 61 texture classes and each class has 92 images captured at different viewpoints and illuminations.
Graph Based Structure Binary Pattern for Face Analysis
2021, OptikCitation Excerpt :The NCDB-LBP solidity is catching of discriminative features from 3 × 3 window by introducing the 4 labeled function. Roy et al. [10] introduced the novel approach by employing the new descriptor called Local Directional ZigZag pattern (LDZP). Initially Local Directional Edge Maps (LDEM) are produced by utilizing kirsch compass masks.
A survey on rotation invariance of orthogonal moments and transforms
2021, Signal ProcessingCitation Excerpt :Their results are based on CASIA NIR and PolyU NIR face databases. Roy et al. [36] demonstrate that the local ZMs-based descriptors outperform many state-of-the-art descriptors for texture classification and human face recognition. Recently, Aggarwal et al. [37] have shown that the OFMMs-based ORIMs descriptors perform much better than the existing state-of-the-art approaches in the retrieval and indexing of medical images.