Texture classification with modified rotation invariant local binary pattern and gradient boosting

Abstract

Since texture is prominent low level feature of an image, most of the image processing and computer vision applications rely on this feature for efficient extraction, retrieval, visualization and classification of the images. Hence, the texture analysis method mainly concentrates on efficient feature extraction and representation of the image. The images captured and analyzed in many of the applications are not in same (or) similar scale, orientation and illumination and also texture has regular, stochastic, periodic, homogeneous (or) inhomogeneous and directional in nature. To address these issues, recent texture analysis method focused on efficient and invariant feature extraction and representation with reduced dimension. Hence this paper proposes a invariant texture descriptor, Locality preserving Rotation Invariant Modified Directional Local Binary Pattern (LRIMDLBP) based on LBP. The classical LBP descriptor is widely used in most of the texture analysis applications due to its simplicity and robustness to illumination changes. However, it does not efficiently capture the discriminative texture information because it uses sign information and ignores the magnitude value of the neighborhood and also suffers from high dimensionality. Hence to improve the performance of LBP, many variants are proposed. Though most of these LBP variants are either geometrical or direction invariant, fails to address the spatial locality and contrast invariance. To address these issues, the proposed LRIMDLBP incorporates spatial locality, contrast and direction information for rotation invariant texture descriptor with reduced dimension. The proposed LRIMDLBP consists of 5 phases: (i) Reference point identification, (ii) Magnitude calculation, (iii) Binary Label computation based on threshold, (iv) Pattern identification in dominant direction and (v) LRIMDLBP code computation. The locality and rotation invariance is achieved by identifying and using reference point in a local neighborhood. The reference point is a dominant pixel whose magnitude is large in the neighborhood excluding center pixel. The spatial locality and rotation invariance is achieved by preserving the weights of LBP dynamically based on the reference point. The proposed method also preserves the direction information of the texture by comparing the magnitude of the pixel in the four dominant directions such as horizontal, vertical, diagonal and anti-diagonal directions. Finally the proposed invariant LRIMDLBP descriptor computes histogram based on decimal pattern value. The proposed LRIMDLBP descriptor results in texture feature with reduced dimension when compared to other LBP variants. The performance of the proposed descriptor is evaluated with large and well known four bench mark texture datasets namely (i) CUReT, (ii) Outex, (iii) KTS-TIPS and (iv) UIUC against three classifiers such as (i). K-Nearest Neighbor (K-NN), (ii). Support Vector Machine (SVM) with Radial Basis Function (RBF) and (iii). Gradient Boosting Classifier (GBC). The intensive experimental result shows that the ensemble based GBC yields superior classification accuracy of 99.38%, 99.43%, 98.67% and 98.82% for the datasets CUReT, Outex, KTH-TIPS and UIUC respectively when compared with other two classifiers and also improves the generalization ability. The proposed LRIMDLBP descriptor achieves approximately 15% more classification accuracy when compared with traditional LBP and also produces 1% to 2.5% more classification accuracy compared with other state of the art LBP variants.

Keywords

Contrast dominant direction ensemble learning gradient boosting LBP Locality rotation invariant texture descriptor

1. Introduction

Texture is an intrinsic property of most of the real world objects and it plays a vital role in many applications such as finger print identification and matching, face and facial expression recognition, object tracking, biodiversity information system, digital library, crime prevention, medical imaging, remote sensing, scene recognition, historical archives and image analysis and recognition. There are numerous methods exist for texture feature extraction and analysis. Th ey can be classified into four categories: (i) Structural (ii) Statistical, (iii) Transform and (iv) Model based methods.

Structural methods are characterized by a set of well defined primitives (texels) and the spatial arrangement of those primitives. Since the performance of this method mainly relies on identification of texels and it is mostly suitable for artificial texture than natural texture. Statistical approach models an image by measuring the spatial distribution and relationship of pixel values. For texture feature extraction and discrimination, either they use simple (first order) or higher order (second and third) statistics. Typical statistical measures of a texture include the gray level co-occurrence matrix [1], grey level difference matrix and run length features [2]. It is less suited for large applications because the computational complexity increases exponentially with the order of statistics. The transform or signal processing method is based on spatial or frequency domain filtering. Generally, Multi-channel filtering is employed to obtain localized spatial-frequency information of a texture. The drawback of this method is number of distinct texture classes should be known in advance.

Model based texture analysis attempt to construct a generative and stochastic image model for estimating the texture feature. The methods reported in the literature are Markov Random Fields (MRF) [3], fractals [4] and multi-resolution autoregressive features [5]. The computational complexity arising in the estimation of stochastic model parameters is the main cause. Fractal model is well suited for natural textures. The main drawback of this method is lack of orientation and hence it is not suitable for local texture structures. Though there are plenty of methods reported in the literature for extracting the texture, in which Local Binary Pattern (LBP) plays an important one in spatial domain.

However, it also has some drawbacks: (i) sensitive to geometric transformation, (ii) High dimension, (iii) Less semantic description of patterns, (iv) missing of spatial encoding among patterns, (v) sensitive to noise and (vi) Magnitude of gray level information is lost. So in order to overcome these limitations, this paper proposes a novel robust Locality preserving Rotation Invariant Modified Directional Local Binary Pattern (RIMDLBP) for Texture Classification. The main contribution of the proposed method is as follows:

(i)
Locality and contrast preserving rotation invariant modified LBP pattern is proposed using the position and magnitude information of the neighboring pixels.
(ii)
The directionality of the texture is identified by sampling the pair of the pixel values in the four dominant direction such as ${0^{\circ}}$ , ${45^{\circ}}$ , ${90^{\circ}}$ and ${135^{\circ}}$ degrees.
(iii)
Both uniform and non uniform patterns are considered to identify the micro texture with reduced dimension.

The rest of the paper is organized as follows: Section 2 presents the related work and Section 3 describes the original Local Binary Pattern (LBP). The various steps involved in the proposed Locality preserving Rotation Invariant Modified Directional Local Binary Pattern (LRIMDLBP) are given in Section 4. The performance measures and the experiments and results are discussed in Section 5. Section 6 describes the conclusion.
2. Related work

The LBP, initially described in [6] and it has many advantages: (i) it is one of the powerful low level feature discriminator with less computational cost, (ii) robust against image intensity changes and (iii) easy to implement. Due to its simplicity and excellent performance, it becomes a reputable texture descriptor in spatial domain. However, it also has some drawbacks: (i) sensitive to geometric transformation, (ii) High dimension, (iii) Less semantic description of patterns, (iv) missing of structural information, (v) sensitive to noise and (vi) loss of magnitude of gray level information. In order to overcome these limitations, variants of LBP methods have proposed in the literature and some of them are discussed below:

Liao et al. [7] have introduced two set of features for texture classification: (i) Dominant Local Binary Pattern (DLBP) and (ii) Circular Symmetric Gabor filter. Since these features are rotation invariant and somehow insensitive tohistogram equalization and noise,it does not capture the distant pixel interaction. Tan and Triggs [8] have reported a more discriminant texture descriptor called Local Ternary Pattern (LTP), which employed three level quantization instead of binary quantization in traditional LBP. It has been proved that the three level quantization processes reduce the noise sensitivity in uniform regions.In order to improve the performance of the LTP, Ji et al. [9] have introduced a multi-scale classification method using a group of Median Local Ternary Pattern (MLTP) descriptors for noisy images. This method consists of three descriptors namely (i) MLTP Central descriptor (MLTP_C), (ii) Radial descriptor (MLTP_R) and (iii) Magnitude descriptor (MLTP_M). These three features are concatenated to form a rotation invariant multi-scale texture feature under different sampling scales and rotation.

The completed LBP (CLBP), which utilizes the both magnitude and sign of gray level differences in a local region havebeen described in [10]. In this method three patterns are introduced namely CLBP_Center (CLBP_C), CLBP_Sign (CLBP_S), CLBP_Magnitude (CLBP_M). These three patterns are fused as a feature vector for texture classification. Though it overcomes the illumination variation in ambient lighting, it produces huge number of features. To address the inadequacy of CLBP, recently Multi Quantized Local Binary Patterns (MQLBP) has been proposed in [11]. This method has higher discrimination power, improved noise robustness and better generalization capability.

The rotation invariance sorted consecutive LBP (scLBP), which builds patterns irrespective of their spatial transition has described in [12]. This method encodes all binary patterns with any number of spatial transitions without scarifying the rotation invariance by using the combination of scLBP and kd-tree. The spatial arrangement of local structure of the texture is considered in [13]. This method utilizes the local spatial information of the image while computing the LBP. In this method the scale parameter and the threshold value of the LBP code are determined using bilateral multi scale filter. Yuan et al. [14] have introduced a joint histogram of LBP and Hamming-Distance-based Local Binary Patterns (HDLBP) to represent the LBP co-occurrence with HDLBP (LBPCoHDLBP) for capturing the spatial structure of the texture.

A rotation invariant Dominant Rotated Local Binary Pattern (DRLBP) has been reported in [15]. In this method, rotation invariance is computed by identifying the reference pixel in a local neighborhood. This descriptor possess both structural and magnitude information of the texture, thereby achieving more discriminative power. A hybrid scheme based on globally rotation invariant matching with locally variant LBP texture features called LBP Variance (LBPV) has been presented in [16]. This method does not have quantization process and also reduces the dimensionality of the feature using distance measurement. Hence they have claimed that, this method achieves improved speed of computation and better classification accuracy over other classical locally invariant LBP methods. These methods are robust to specific rotation angles only. This limitation is addressed in [17]. The Principal Curvatures based LBP (PCLBP) descriptor is constructed by employing the complementary information of LBP and Principal Curvatures (PCs) of texture surface. The PCs, which consist of minimum and maximum curvatures, can capture the macro and microstructure texture information and have the capability of consecutive rotation invariance. The PCLBP is rotation and illumination invariant descriptor with less computational cost.

The Local Directional Ternary Pattern (LDTP) has been introduced in [18]. This method combines contrast and directional information from LTP and LDP descriptors respectively. To achieve robustness and get more detailed information, this method uses Frei-Chen masks and second order Gaussian filter over the (3 $\times$ 3) overlapping windows of the entire image. Hence it suffers from high computational complexity and dimensionality. Nguyen et al. [19] have proposed a method called rotation invariant Statistical Binary Patterns (SBP). In this method, the rotation invariant LBP operators are applied to series of moment images. Since this method is robust to noise, it suffers from high computational cost and limited discriminative power due to the higher proportion of non uniform patterns. By considering the gradient magnitude and orientation, an adaptive scale and orientation LBP code has been computed in [20]. In addition a weighted LBP feature based on the minimum magnitude of local differences has also been utilized. A rotation invariant Multiscale Co-occurrence LBP (MCLBP) for texture classification has introduced in [21]. The authors have claimed that, this method considers the correlation among different scales around the center pixel, which is missing in Multiscale LBP (MLBP).

Figure 1.

LBP Computation. (a) (3 $\times$ 3) window (b) Threshold (c) Weights (d) LBP binary pattern (e) Decimal value.

Center Symmetric Local Binary Co-occurrence pattern for texture has been presented in [22]. In this method, the LBP code is constructed by considering different directions and distances. It obtains texture feature from the co-occurrence of LBP codes, while the traditional LBP extracts frequency information from histogram. Merabet et al. [23] have introduced a variant of Center Symmetric LBP (CS-LBP) called Attractive CS-LBP (ACS-LBP) and Repulsive CS-LBP (RCS-LBP). These patterns consider the four triplets corresponding to the vertical and horizontal directions, and the two diagonal directions by including the value of the central pixel. In addition to that, Average Local Gray Level (ALGL), Average Global Gray Level (AGGL) and the median value over (3 $\times$ 3) neighborhood are introduced to capture both microstructure and macrostructure texture information.

A scale and orientation adaptive extension of Local Binary Pattern has been described in [24]. This method has high computational complexity as compared with lightweight LBPs and produces discriminative and reliable features for noisy images. A new rotation invariant Feature based Local Binary Pattern (FbLBP) have described in [25]. In this method, difference vector is decomposed into sign and magnitude part, the sign part is described by conventional LBP, while the magnitude part is described by two features of the mean and the variance of the magnitude vector. They claimed that these two features have high complementarily to the sign part and sensitive to illumination changes with a low dimensionality. To obtain statistical, directional and structural characteristics of the texture, recently wavelet based LBP methods are introduced in [26, 27, 28, 29].

Generally the texture has regular, stochastic, periodic, homogeneous (or) inhomogeneous and directional in nature and the images used for training and testing are not captured in same (or) similar scale, orientation and illumination. Since the LBP and its variants are excellent in capturing the structure of the image in spatial domain, but the state of the art LBP based texture descriptor discussed above is robust to either one (i.e. Rotation invariant) or the combination of one or more issues (i.e. rotation invariant with illumination, scale, spatial and direction invariant) resulting a feature descriptor in high dimension. Hence in order to improve the classification accuracy, this paper proposes a robust Locality preserving Rotation Invariant Modified Directional Local Binary Pattern (RIMDLBP) for Texture Classification. The classical LBP and the proposed RIMDLBP are described in forthcoming sections.

3. Local binary pattern (LBP)

Ojala et al. [6] have proposed the Local Binary Pattern (LBP) to represent the local gray level structure. The LBP considers the local neighborhood around each pixel and compare the neighborhood pixel with center pixel. The neighboring pixels are assigned a binary label, which can be either 0 or 1 depending on whether the center pixel has higher intensity value than the neighboring pixel. These binary values are multiplied by specific weights and summed up. The LBP code is given by

$\displaystyle\textit{LBP}_{P,R}\left({{x_{c}},{y_{c}}}\right)=\sum\limits_{P=0% }^{P-1}{s\left({f\left({{x_{p}},{y_{p}}}\right)}\right.}\left.{-f\left({{x_{c}% },{y_{c}}}\right)}\right){2^{p}}$ (1)

where $s(z)$ is the threshold function based on sign information and $z=\left({f\left({{x_{p}},{y_{p}}}\right)-f\left({{x_{c}},{y_{c}}}\right)}\right)$ The $s(z)$ is defined as

$\displaystyle s\left(z\right)=\left\{{\begin{array}[]{*{20}{ll}}1&{z\geqslant 0% }\\ 0&\text{otherwise}\end{array}}\right.$

where $f\left({{x_{c}},{y_{c}}}\right)$ and $f\left({{x_{p}},{y_{p}}}\right)$ are the center and neighborhood pixel values respectively and $P$ stands for the number of pixels in the neighborhood and $R$ is the radius of the neighborhood which determines how far the neighboring pixels are located away from the center pixel. The pixel values of the neighbors which are not in the image grids can be calculated using an interpolation method. The illustration of LBP computation is shown in the Fig. 1.

A binary pattern is computed for all the pixels in the image and is called the local binary map of the image. Then the histogram $H{\left(L\right)_{\text{LBP}}}\left({L=0,1,2,\ldots,{2^{P}}\!-\!1}\right)$ of this local binary map can be created as a texture feature vector and is represented in the Eq. (2) as follows:

$\displaystyle H{\left(L\right)_{\textit{LBP}}}=\frac{{\sum\limits_{{x_{c}}=0}^% {u-1}{\sum\limits_{{y_{c}}=0}^{v-1}{f\left({{x_{c}},{y_{c}}}\right)}}}}{{u% \times v}}$ (2)

where

$\displaystyle f\left({{x_{c}},{y_{c}}}\right)=\left\{\begin{array}[]{ll}1&% \text{if LBP}\left({{x_{c}},{y_{c}}}\right)=h\\ 0&\text{otherwise and}\ {u,v}\ \text{are the size}\\ &\quad\text{of the image}.\end{array}\right.$

To obtain the rotation invariant LBP, Ojala et al. [30] have presented the following steps as:

$\displaystyle\textit{LBP}_{P,R}^{\textit{riu}\beta}=$ $\displaystyle\quad\left\{\begin{array}[]{ll}\sum\limits_{p=0}^{p-1}s(f(x_{p},y% _{\rho})&\text{if}\ U\left(\textit{LBP}_{P,R}\right)\leqslant 2\\ \quad-f(x_{c},y_{c}))\\ P+1&\text{otherwise}\end{array}\right.$ (3)

where ${U\left({\textit{LBP}_{P,R}}\right)}$ is a Uniform LBP and is defined as:

$\displaystyle U\left({\textit{LBP}_{P,R}}\right)=\left|{s\left({f\left({{x_{p}% },{y_{p}}}\right)-f\left({{x_{c}},{y_{c}}}\right)}\right)}\right.\left.{-s% \left({f\left({{x_{1}},{y_{1}}}\right)-f\left({{x_{c}},{y_{c}}}\right)}\right)% }\right|+\sum\limits_{i=2}^{P-1}{\left|s\right.}\left({f\left({{x_{i}},{y_{i}}% }\right)-f\left({{x_{c}},{y_{c}}}\right)}\right)\left.{-s\left({f\left({{x_{i-% 1}},{y_{i-1}}}\right)-f\left({{x_{c}},{y_{c}}}\right)}\right)}\right|$

Though the LBP descriptor is robust to rotation invariant, it fails to capture the locality, direction and contrast information. Hence the proposed method overcomes the limitations of traditional LBP and the details are presented in the next section.

4. Proposed locality preserving rotation invariant modified directional local binary pattern (LRIMDLBP)

The classical LBP descriptor uses only the sign information to compute the descriptor and ignores the magnitude value. In order to increase the discriminative power of the descriptor, the magnitude information is used in [9] and the LBP is rotation variant due to fixed arrangement of weights. Though, the rotation invariant ${\textit{LBP}_{P,R}^{\textit{riu}\beta}}$ descriptor performs well, but it has a drawback due to Eq. (1): (i) Losing locality and (ii) unable to describe how much change exists between center pixel and neighbors. For example assume two situations:

Case 1:
Consider two textons whose center pixels are 63 and the neighboring pixels are {14, 52, 58, 70, 78, 53, 221, 43} and {22, 42, 50, 80, 92, 58, 250, 52} respectively. The LBP codes are same for both the textons as {1, 1, 0, 1, 0, 0, 0, 0}. Though the LBP codes are same for these two textons, their contrast (magnitude) (i.e. the absolute difference between neighboring pixel and center pixel) information is different whose values are {49, 11, 5, 7, 15, 10, 158, 20} and {41, 21, 13, 17, 29, 5, 187, 11} respectively. So the traditional LBP ( ${\textit{LBP}_{P,R}^{\textit{riu}\beta}}$ ) cannot accommodate the contrast change while coding.
Case 2:
Consider two textons whose center pixels are 50 and the neighboring pixels are {14, 52, 58, 70, 78, 53, 221, 43} and {58, 70, 78, 53, 221, 43, 14, 52} respectively. The LBP code for these textons are {1, 1, 0, 1, 0, 0, 0, 0} and {0, 1, 0, 0, 0, 0, 1, 1} respectively. It is observed that, the pixels in the textons are same but their positions are different. Since the neighborhood pixels are same, their relative position changes resulting in two different LBP code. This is because the LBP samples the neighboring pixels in a static order. Hence it is inferred that the ${\textit{LBP}_{P,R}^{\textit{riu}\beta}}$ does not preserve the spatial locality of the pixel while coding.

Figure 2.
Framework for the proposed LRIMDLBP computation.

Figure 3.
Proposed LRIMDLBP computation. (a) Example (3 $\times$ 3) window (b) reference point identification (c) Magnitude computation (d) Binary Label computation based on $t$ (e) pattern identification in dominant direction (e) Weights Assignment based on reference point (f) proposed LBP binary pattern (g) Decimal value.

Based on the two observation it is identified that, the ${\textit{LBP}_{P,R}^{\textit{riu}\beta}}$ unable to code the contrast and locality information which is essential to describe the texture. So this paper proposes the contrast and locality preserving rotation invariant descriptor to supplement these missing information besides traditional LBP. The block diagram of the proposed method is shown in Fig. 2.

Initially the input image is converted into gray scale and is given to Gaussian filtering to reduce the noise and complicated patterns. Then it is normalized using Z-score normalization to reduce illumination variation. The LRIMDLBP pattern is extracted and the decimal equivalent value is computed. Histogram is constructed based on the decimal value and then these values are classified with Gradient boost classifier.

The proposed LRIMDLBP utilizes magnitude information to derive the rotation invariant texture feature descriptor. The proposed LRIMDLBP consists of 5 phases: (i) Reference point identification, (ii) Magnitude calculation, (iii) Binary Label computation based on threshold, (iv) Pattern identification in dominant direction and (v) LRIMDLBP code computation. In the reference point identification phase, the dominant pixel value is identified from the ( $n\times n$ ) neighborhood and the weights are arranged in a clockwise manner based on the dominant pixel. The dominant pixel whose magnitude is maximum in the ( $n\times n$ ) neighborhood excluding center pixel and is represented in Eq. (4) as:

$\displaystyle{D_{P}}=\arg\max\mathop{\left({f\left({{x_{p}},{y_{p}}}\right)}% \right)}\limits_{p\in 0..p-1}$ (4)

The dominant pixel ( ${D_{P}}$ ) is called as reference point from which the pixel positions are arranged dynamically in a clock wise manner. There is an issue addressed here.
4.1 More than one pixel as dominant pixel

In this case, the pixel values in a ( $n\times n$ ) neighborhood except the central pixel is subtracted from the weight value of the pixel (i.e. ${2^{0}}$ , ${2^{1}}$ , $\ldots$ , ${2^{8}}$ ) with respect to the central pixel in a clock wise manner in horizontal direction. Then the resultant block is used for further analysis and dominant pixel identification. For example the dominant pixel ${D_{P}}$ is in ${P^{th}}$ position, $P\in 0$ .. $P-1$ then the dynamic sampling order of the pixel positions are:

$\displaystyle\text{if}\ \textit{oldp}=\left({p+i}\right)\ \text{mod}\ 8\ \text% {then}\ \textit{newp}=i,i\in 0..p-1$ (5)

The proposed method uses an adaptive sampling order based on reference point ${D_{P}}$ to achieve the spatial locality preserving and rotation invariance LBP. In the second phase, initially the magnitude which represents the contrast change between the center and the neighboring pixels in a block is computed. The magnitude ${M_{P}}$ is the absolute difference between the center pixel and the neighboring pixel value and is represented in Eq. (6).

$\displaystyle{M_{p}}=\left|{f\left({{x_{c}},{y_{c}}}\right)\mathop{-}\limits_{% p=0}^{p-1}f\left({{x_{p}},{y_{p}}}\right)}\right|$ (6)

Then the proposed method identifies the binary label (BL) for each neighboring pixel in order to compute the pattern as follows:

$\displaystyle B{L_{i}}=\mathop{b\left({{x_{i}}}\right)}\limits_{i=0..p-1}$ (7)

$\displaystyle b\left({{x_{i}}}\right)=\left\{{\begin{array}[]{*{20}{c}}1&{% \text{if}\ {M_{p}}\leqslant t}\\ 0&{\text{otherwise}}\end{array}}\right.$

Figure 4.

Proposed rotation invariant LRIMDLBP computation. (a) (3 $\times$ 3) window (b) ${90^{\circ}}$ rotation of (a). (c) reference point identification (d) Magnitude computation (e) Binary Label computation based on $t$ (f) pattern identification in dominant direction (g) Weights Assignment based on reference point (h) proposed LBP binary pattern (i) Decimal value.

where $t$ is the threshold, which is the median of the absolute local differences between neighboring pixels and center pixel in the block. The threshold $t$ quantizes the contrast change and also determines the uniformity of the pixel in the block. Generally, the binary values are ordered in a clock wise manner to obtain the pattern. Since the ordering is performed statically, it leads to different pattern. Hence, the proposed method computes the locality and contrast preserving rotation invariant texture descriptor by sampling the neighborhood in an adaptive manner using reference point ${D_{P}}$ . The weights are also arranged in the neighborhoods horizontally in a clockwise manner from ${D_{P}}$ . The traditional LBP and LTP [8] creates the patterns circularly and considers the uniform relationship between the neighboring pixels whereas the proposed method encodes the pattern by considering the spatial relationship between any pair of neighbors in the given block along with the dominant directions. The proposed LRIMDLBP catches the direction by comparing the binary labels in four directions such as (i) horizontal (ii) vertical, (iii) diagonal and (iv) anti diagonal with respect to the angles such as ${0^{\circ}}$ , ${45^{\circ}}$ , ${90^{\circ}}$ and ${135^{\circ}}$ respectively. The binary label (BL) in these directions is compared with respect to the Eq. (4.1) and is defined as:

$\displaystyle\textit{BL}=$ (8) $\displaystyle\quad\left\{\begin{array}[]{ll}1&\text{if}\left(\left(\textit{BL}% _{0}\vee\textit{BL}_{4}\right)=1\right)\ \theta={0^{\circ}},\\ &\quad\left(\left(\textit{BL}_{1}\vee\textit{BL}_{5}\right)=1\right)\ \theta=4% 5^{\circ}\\ 0&\text{otherwise}\end{array}\right.$

The decimal pattern for the block is obtained by multiplying the binary values obtained using the Eq. (4.1) with the weights. The weights are arranged based on the reference point ${D_{P}}$ . The diagrammatic representation of the proposed method with rotation invariant is shown in Figs 3 and 4 respectively. The above mentioned process is repeated for all the blocks and obtains the decimal pattern value. As with traditional LBP, the histogram is computed based on the decimal pattern value for texture feature. Though the proposed method follows the same process of LBP histogram as a texture feature, the dimensionality is greatly reduced and also concentrates on directionality along with magnitude and sign. The steps are represented as pseudo code in Fig. 5. The next section describes the experimental result and the performance measure used in this paper for evaluating the proposed and other state of the art methods.

Figure 5.

Pseudo code of the proposed LRIMDLBP Method.

5. Experiments and results

The proposed method LRIMDLBP is experimented with four dataset such as CUReT [31], Outex [32], KTH-TIPS [33] and UIUC [34]. It is implemented in windows environment using JDK10 and Scikit learn. The performance of the proposed method is measured with classification accuracy which is defined as number of images correctly classified by total number of images and is shown below:

$\displaystyle\textit{classification Accuracy}=\frac{c}{s}\cdot 100$ (9)

where $c$ is number of images correctly classified and $s$ is the total number of images in the sample.

During experimentation, the proposed method converts the input color image into gray scale image. In order to remove noise and retain the smoothness, the Gaussian filter is applied with the kernel size $k$ as 5. The smoothed image is divided into overlapping blocks and for each block the proposed rotation invariant LRIMDLBP code is extracted. Initially, the reference point ${D_{P}}$ is computed with the Eq. (4) from each block.Based on ${D_{P}}$ the position of the neighboring pixelsare reordered. The process of reordering is given in step 3 of Fig. 5. The magnitude ${M_{P}}$ is calculated with Eq. (6) for the neighboring pixels, which represents the contrast (or) intensity variation. The proposed method then identifies the appropriate threshold value $(t)$ in order to compute binary label for each block. The process of threshold calculation is presented in step 5 in Fig. 5. Based on this threshold $(t)$ , the binary label is obtained using the Eq. (7) as represented in the Section 4. Then the rotation invariant directional pattern in dominant direction is identified with the Eq. (4.1). Finally, the proposed rotation invariant LRIMDLBP pattern is computed based on step 8 of Fig. 5. The same process is repeated for each block and the proposed LRIMDLBP code is obtained for the given image. With these code, the histogram is computed with 16 bins and these bins are normalized with the Eq. (10) as:

$\displaystyle{H_{nor}}=\frac{{n_{k}}}{\textit{size of the image}}$ (10)

where ${n_{k}}$ is the value of the histogram and ${H_{nor}}$ is the normalized histogram. The normalized values are stored as a feature for each image. The above mentioned process is repeated for all the images in the dataset and the corresponding feature vector is stored in the feature database. The histogram is considered as feature vector and is modeled by the classifier to obtain the accuracy of the classification. During classification process, the proposed method splits the feature dataset in the ratio of (0.8, 0.2) for training and testing respectively with 5 fold cross validation.

The performance of the proposed method is measured with three different classifiers such as (i) k-Nearest Neighbor (k-NN), (ii) Support Vector Machine (SVM) and (iii) Ensemble classifier. The proposed method adapts a class of ensemble learning based on Gradient Boosting classifier. In boosting method, the base estimators are combined to produce a powerful ensemble which will address the bias issue effectively. Generally the accuracy of the classifier depends on parameters of the algorithm. The proposed method is implemented with four different dataset with three different classifiers. The same parameter set for a particular classifier is not suitable for all the four dataset. Hence the proposed method identifies the appropriate hyper parameter for each classifier with grid search method and to measure the accuracy of the model with 5 fold cross validation. The hyper parameter set used for tuning for each classifier is presented in Table 1.

Table 1

Parameter set for k-NN, SVM and Gradient Boost classifiers for tuning

Name of the classifier	Parameter set used for tuning
k-NN	Neighbors(k): $\left\{{[1,3,5,7,10]}\right\}$
	Weights {uniform}
	Distance metric(d): {Euclidean, Manhattan}
SVM	Kernel: Radial Basis Function (RBF)
	Gamma: $\left\{{1,0.1,0.01,0.001}\right\}$ ,
	Regularization (C): $\left\{{0.1,1,10,100}\right\}$
Ensemble Classifier (Gradient Boost)	Learning rate: $\left\{{0.01,0.025,0.05,0.075,0.1,0.15,0.2}\right\}$
	Max_depth : $\left\{{3,5,6,7,8}\right\}$
	Subsample: $\left\{{0.6,0.7,0.8}\right\}$
	n_estimators: $\left\{{100,250,500,750,1000}\right\}$

Figure 6.

Classification accuracy of the three classifiers for the proposed LRIMDLBP method.

The classification accuracy of the classifiers for the four dataset is obtained and is shown in the Fig. 6. From the figure, it is evident that the gradient boosting performs well when compared to other classifiers and yields high classification accuracy for all the four dataset. The k-NN produces high classification accuracy such as 97.52%, 98.56% and 97.51% for CUReT, KTH-TIPS and UIUC dataset respectively when compared with SVM. The SVM and gradient boosting gives similar accuracy for the Outex and UIUC datasets.

The experiments are conducted with different neighborhood and radius values. The P, R values considered in this experiments are (8, 1) , (16, 2) and (24, 3). During experimentation it is observed that, the higher values of P, R yields less classification accuracy when compared with P $=$ 8 and R $=$ 1 for the proposed method. The proposed method is also compared with $\textit{LBP}^{\textit{riu}}$ [30], DRLBP [15], LDTP [18], ScLBP [12], FbLBP [25], LBPV [16] and MLTP [9] methods by taking the experimental values cited in the paper. The experimental results for the proposed and other methods for each dataset are discussed below:

5.1 CUReT

The Columbia-Utrecht (CUReT) [31] database contains 61 surface texture classes and each class posses 205 images with different illumination and viewing angles. Among 205 images from each class, we have considered only 92 images with less than ${60^{\circ}}$ viewing angle. In the proposed work, each image is cropped into (200 $\times$ 200) across all classes and is converted into gray scale.

The experiments are conducted with different number of images in the training set. The number of images considered for training is 46, 26, 12 and 6 and the remaining images are for testing. The value of $\sigma$ , $k$ is set to 0.5 and 5 respectively for Gaussian filter. The experiments are conducted with these values and the classification accuracy of 99.38%, 97.62%, 95.74% and 88.56% for the training images of 46, 26, 12 and 6 respectively are obtained. From the experimental result it is evident that, when the number of training images is increased, the performance is also increased. Hence the proposed method uses 46 images as training images (i.e 50% images)with the gradient boost classifier and the obtained result is presented in Table 2. The experiments are conducted in order to compare the performance of the proposed LRIMDLBP method with other existing methods as mentioned above for various values of P,R. The obtained results are incorporated in Table 2. It is noted from the Table 2 that the ScLBP and FbLBP methods yield the classification accuracy of 98.11% and 97.28% respectively. It is evident from the Table 2 that the proposed method outperforms with these two methods. It is also observed that, the proposed and FbLBP method give the accuracy for the value of P $=$ 8, R $=$ 1, which automatically reduces the dimensionality and computational complexity. The classical $\text{LBP}^{\textit{riu}2}$ whose dimensionality is less (10 only) when compared to the proposed and other methods, but fails to produce high classification accuracy because it considers the sign information and uniform patterns. The proposed method achieves high accuracy shown by the method scLBP over 1.27%,

Table 2
Classification accuracy of the proposed and other methods for CUReT database

	P $=$ 8, R $=$ 1	P $=$ 16, R $=$ 2	P $=$ 24, R $=$ 3
LRIMDLBP	99.38	98.33	96.04
$\text{LBP}^{\textit{riu}2}$	67.23	67.02	67.52
LBPV	62.73	56.23	56.65
MLTP	96.23	95.23	94.62
LDTP	92.54	–	–
DRLBP	95.59	95.05	94.27
ScLBP	97.41	98.11	97.92
FbLBP	97.28	97.12	97.04
BRINT	95.02	93.95	94.12

5.2 Outex

The Outex [32] database contains 24 different classes of texture images which are captured under three different illuminations (“horizon”, “inca” and “t184”) and various rotation angles such as ( ${0^{\circ}}$ , ${5^{\circ}}$ , ${10^{\circ}}$ , ${15^{\circ}}$ , ${30^{\circ}}$ , ${45^{\circ}}$ , ${60^{\circ}}$ , ${75^{\circ}}$ , and ${90^{\circ}}$ ). From each image, twenty non-overlapping texture samples with window size (128 $\times$ 128) are obtained. In this experiment TC10 and TC12 suites are considered. TC10 uses “inca” illuminations and orientation 0 of each class for training and the other 8 orientation angles with the same illuminations for testing. So, there are 480 (24 $\times$ 20) training samples and 3840 (24 $\times$ 20 $\times$ 8) testing samples are used for conducting experiments. TC12 uses the same training samples as TC10 and uses all samples captured under illuminations “t184” or “horizon” for testing. Hence, there are 480 (24 $\times$ 20) training samples and 4320 (24 $\times$ 20 $\times$ 9) testing samples for each illumination.

With these training images, the experiments are conducted based on the steps explained above. The classification accuracy of the proposed LRIMDLBP is presented in Table 3. The proposed method yields the classification result for the TC-10 and TC-12 dataset as 99.43% and 97.24% respectively for P $=$ 8, R $=$ 1. The classical $\text{LBP}^{\textit{riu}2}$ method also performs equally well when compared to other methods for the data set TC10. The MLTP and FbLBP produce the classification accuracy of 99.48% and 99.58% respectively. They differ in accuracy with the proposed LRIMDLBP by 0.05 and 0.15. For the TC10 dataset LBPV performs better when compared with $\text{LBP}^{\textit{riu}2}$ , DRLBP, ScLBP and BRINT. The proposed LRIMDLBP gives high classification accuracy, which is 98.26% among all the methods in TC12 dataset. It implies that the proposed method is more robust to illumination changes than other methods. The DRLBP, ScLBP and FbLBP also yields better result with P $=$ 24 and R $=$ 3 whereas BRINT with P $=$ 16, R $=$ 2. From these discussion, it is inferred that the proposed method capture the structure of the micro texture very well and also robust to illumination changes with less dimension.

Table 3
Classification result of the proposed and other methods for TC-10 and TC-12

	P $=$ 8, R $=$ 1		P $=$ 16, R $=$ 2		P $=$ 24, R $=$ 3
	TC10	TC12	TC10	TC12	TC10	TC12
LRIMDLBP	99.43	97.59	98.26	95.65	97.22	95.11
$\text{LBP}^{\textit{riu}2}$	84.87	71.36	89.40	82.39	95.16	87.03
LBPV	96.26	84.92	97.02	89.38	98.15	89.79
MLTP	99.48	97.15	98.23	96.64	98.52	96.44
LDTP	93.52	90.75	–	–	–	–
DRLBP	94.98	94.92	95.23	94.74	96.23	96.52
ScLBP	94.53	93.58	95.49	96.16	97.20	97.26
FbLBP	98.85	93.82	96.64	95.52	99.58	96.82
BRINT	91.72	88.24	96.43	94.59	94.24	94.87

5.3 KTH-TIPS

The KTH-TIPS [33] database is a challenging database, which contains 10 texture classes. The images are captured in nine different scales, three different poses and three illumination conditions resulting in 81 images per class. A random split applied on each class of texture images and obtained 40 images for training and the rest is for testing. Since the KTH-TIPS database is not rotation invariant, but for the proposed method rotates the image into ${45^{\circ}}$ , ${90^{\circ}}$ , ${135^{\circ}}$ and ${180^{\circ}}$ degree rotations for experiment. The Table 4 shows the experimental result obtained for the proposed LRIMDLBP and other methods in different scales. The proposed method achieves 98.67% classification accuracy which is superior with the results obtained by other methods. KTH-TIPS is a complex dataset with more intra class variation apart from rotation, illumination and scale changes. However the proposed LRIMDLBP still yields better results compared with other methods and proves its robustness against rotation variation and illumination. The MLTP and BRINT perform well while compared with other methods. It is noted that, LRIMDLBP, MLTP and BRINT achieves better result for the scale P $=$ 8 and R $=$ 1. Hence it is inferred that the small neighborhood captures the spatial texture information efficiently even for the complex dataset than with the larger neighborhood.

Table 4
Classification accuracy of proposed and other methods for KTH-TIPS dataset

	P $=$ 8, R $=$ 1	P $=$ 16, R $=$ 2	P $=$ 24, R $=$ 3
LRIMDLBP	98.67	96.22	94.21
$\text{LBP}^{\textit{riu}2}$	83.65	88.29	88.53
LBPV	78.98	82.76	81.01
MLTP	98.25	98.22	97.25
LDTP	90.52	–	–
DRLBP	93.17	94.80	90.16
ScLBP	78.39	77.52	77.02
FbLBP	92.32	92.01	90.23
BRINT	97.28	97.21	96.22

5.4 UIUC

The UIUC [34] dataset has 25 classes of texture images, with each class containing 40 images. The database images undergone to different scale, rotation and illumination variations. The size of each image is (640 $\times$ 480) and for this experiment, we crop the images into (256 $\times$ 256) pixels and converted to a gray-scale image. The classification result of the proposed LRIMDLBP and other methods are presented in Table 5. The LRIMDLBP slightly performs well when compared with ScLBP for the scale P $=$ 16 and R $=$ 2. MLTP obtains better results than FbLBP and BRINT. Since the FbLBP and BRINT consistently perform well for other dataset, they produce lower accuracy for this dataset. DRLBP is superior to LDP and FbLBP and LDP methods.

Table 5
Classification Accuracy of proposed and other methods for UIUC dataset

	P $=$ 8, R $=$ 1	P $=$ 16, R $=$ 2	P $=$ 24, R $=$ 3
LRIMDLBP	97.85	98.42	96.07
$\text{LBP}^{\textit{riu}2}$	66.23	72.92	72.01
LBPV	66.61	73.31	75.01
MLTP	96.97	97.22	97.52
LDTP	93.23	–	–
DRLBP	95.72	94.63	93.72
ScLBP	98.45	98.21	98.03
FbLBP	92.04	94.14	94.17
BRINT	79.52	84.14	86.39

The dimensionality of the feature plays a vital role in the performance analysis. The methods consider above for performance analysis utilizes the rotation invariant uniform patterns. The experimental study proved that the non uniform patterns also significantly improve the classification accuracy so the proposed method utilizes both uniform and non-uniform patterns. Since the DRLBP [16] used a dictionary learning to select the appropriate patterns for classification, the original feature length is 256. The discriminative patterns are obtained by incorporating kd-Tree approach in scLBP [15] method. So the length of the features is dynamically varying with different dataset. The actual feature length of MLTP is 6561 for P $=$ 8, R $=$ 1, whereas for the rotation invariant mapping of $\text{MLTP}^{\textit{riu}3}$ is 24 only. The dimensionality of the proposed LRIMDLBP and other methods are presented in Table 6. In the proposed method the length of the feature vector is static and the length is 16 (P $=$ 3, R $=$ 1), which is low when compared with other variations of the LBP method. The experimental result proved that the proposed method gives high accuracy with the small sampling neighborhood when compared with other method. Since the dimensionality of classical $\text{LBP}^{\textit{riu}2}$ method is 10 which is very low related with the proposed and other methods but it produces low classification accuracy for all the dataset considered for this experiment. It considers only uniform patterns whereas the proposed method considers all the patterns resulting in more discriminative power.

Table 6

Dimensionality of feature (for P $=$ 8, R $=$ 1) for proposed and other LBP based methods

	Feature length
LRIMDLBP	16
$\text{LBP}^{\textit{riu}2}$	10
LBPV	10
MLTP (original)	6561
$\text{MLTP}^{\textit{riu}3}$	24
LDTP	1022
DRLBP	256
ScLBP	Dynamic
FbLBP	80
BRINT	144

6. Conclusion

In this paper, a novel texture descriptor called as Locality preserving Rotation Invariant Modified Directional Local Binary Pattern (LRIMDLBP) is proposed. The proposed descriptor utilizes the magnitude and adaptive sampling of the order of the pixel neighborhood to achieve contrast and locality preserving LBP. The proposed method also preserves the direction information in four dominant directions such as horizontal, vertical, diagonal and anti-diagonal directions resulting in less dimension of the texture feature. The rotation invariance is achieved by identifying the reference point in the local neighborhood. The experimental analysis of the proposed method is performed on four different dataset such as CUReT, Outex, KTS-TIPS and UIUC with three classifiers namely k-NN, SVM, and gradient boosting. The proposed method obtains better classification accuracy based on gradient boosting classifier when compared with other classical LBP variants. The proposed method is extended for color texture classification and other classification applications in future.

References

Haralick

Shanmugam

Dinstein

. Textural features for image classification. IEEE Trans Syst Man Cybern. 1973; 3(6): 610-621.

Tang

. Texture information in run-length matrices. IEEE Trans Image Process. 1998; 7(11): 1602-1609.

Chen

Huang

. Markov random fields for texture classification. Pattern Recognit Lett. 1993; 14(11): 907-914.

Chen

. On texture classification using fractal dimension. Intern J Pattern Recognit Artif Intell. 1999; 13(6): 929-943.

Mao

Jain

. Texture classification and segmentation using multiresolution simultaneous autoregressive models. Pattern Recognit. 1992 Feb; 25(2): 173-188.

Ojala

Pietikainen

Harwood

. A comparative study of texture measures with classification based on feature distributions. Pattern Recognit. 1996; 19(3): 51-59.

Liao

Law

Chung

. Dominant local binary patterns for texture classification. IEEE Trans Image Process. 2009; 18(5): 1107-1118.

Tan

Triggs

. Enhanced local texture feature sets for face recognition under difficult lighting conditions. IEEE Trans Image Process. 2010; 19(6): 1635-1650.

Ren

Liu

. Median local ternary patterns optimized with rotation-invariant uniform-three mapping for noisy texture classification. Pattern Recognit. 2018 Jul; 79: 387-401.

10.

Guo

Zhang

. A completed modeling of local binary pattern operator for texture classification. IEEE Trans Image Process. 2010; 19(6): 1657-1663.

11.

Patel

Maheshwari

Balasubramanian

. Multi-quantized local binary patterns for facial gender classification. Comput Electr Eng. 2015; 54: 271-284.

12.

Ryu

Hong

Yang

. Sorted consecutive local binary pattern for texture classification. IEEE Trans Image Process. 2015; 24(7): 2254-2265.

13.

Kaddar

Fizazi

Boudraa

. Texture features based on an efficient local binary pattern descriptor. Comput Electr Eng. 2018 Aug; 70: 496-508.

14.

Yuan

Xia

Shi

. Mixed co-occurrence of local binary patterns and hamming distance based local binary patterns. Inf Sci. 2018 Sep; 460461: 202-222.

15.

Mehta

Egiazarian

. Dominant rotated local binary patterns (DRLBP) for texture classification. Pattern Recognit Lett. 2016; 71: 16-22.

16.

Guo

Zhang

. Rotation invariant texture classification using LBP variance (LBPV) with global matching. Pattern Recognit. 2010; 43: 706-719.

17.

Kou

Cheng

. Chen

. Zhuang

. Principal curvatures based local binary pattern for rotation invariant texture classification. Optik. 2019 Sep; 193: 162999.

18.

Khadiri

Chahi

Merabet

Ruichek

Touahni

. Local directional ternary pattern: A new texture descriptor for texture classification. Comput Vis Image Underst. 2018 Apr; 169: 14-27.

19.

Nguyen

Manzanera

. Statistical binary patterns for rotational invariant texture classification. Neurocomputing. 2016 Jan; 173(3): 1565-1577.

20.

Davarzani

Mozaffari

Yaghmaie

. Scale and rotation-invariant texture description with improved local binary pattern features. Signal processing. 2015 Jun; 111: 274-293.

21.

Shen

Zhao

Pietikäinen

. Globally rotation invariant multiscale co-occurrence local binary pattern. Image Vis Comput. 2015 Nov; 43: 16-26.

22.

Verma

Raman

. Center symmetric local binary co-occurrence pattern for texture, face and biomedical image retrieval. J Vis Commun Image Represent. 2015; 32: 224-236.

23.

Merabet

Ruichek

Idrissi

. Attractive and repulsive center-symmetric local binary patterns for texture classification. Eng Appl Artif Intell. 2019 Feb; 78: 158-172.

24.

Hegenbart

Uhl

. A scale and orientation adaptive extension of local binary patterns for texture classification. Pattern Recognit. 2015 Aug; 48(8): 2633-2644.

25.

Pan

Fan

. Feature based local binary pattern for rotation invariant texture classification. Expert Syst Appl. 2017; 88: 238-248.

26.

Nithya

Ramakrishnan

. Wavelet domain directional binary pattern using majority principle for texture classification. Physica A. 2020 May; 545: 123575.

27.

Gopala Krishnan

Vanathi

. An efficient texture classification algorithm using integrated discrete wavelet transform and local binary pattern features. Cogn Syst Res. 2018 Dec; 52: 267-274.

28.

Ramakrishnan

Nithya

. Two improved extension of local binary pattern descriptors using wavelet transform for texture classification. IET Image Process. 2018; 12(11): 2002-2019.

29.

Muqeet

Holambe

. Local binary patterns based on directional wavelet transform for expression and pose-invariant face recognition. Appl Comput Inform. 2019; 15(2): 163-171.

30.

Ojala

Pietikainen

Maenpaa

. Multiresolution gray scale and rotation invariant texture classification with local binary patterns. IEEE Trans Pattern Anal Mach Intell. 2002; 24(7): 971-987.

31.

http://www.cs.columbia.edu/CAVE/software/curet/.

32.

http://www.outex.oulu.fi/index.php?page=classification.

33.

http://www.nada.kth.se/cvap/databases/kth-tips/.

34.

http://www-cvr.ai.uiuc.edu/poncegrp/data/index.html.

Texture classification with modified rotation invariant local binary pattern and gradient boosting

Abstract

Keywords

1. Introduction

Table 2 Classification accuracy of the proposed and other methods for CUReT database

Table 3 Classification result of the proposed and other methods for TC-10 and TC-12

Table 4 Classification accuracy of proposed and other methods for KTH-TIPS dataset

Table 5 Classification Accuracy of proposed and other methods for UIUC dataset

References

Table 2
Classification accuracy of the proposed and other methods for CUReT database

Table 3
Classification result of the proposed and other methods for TC-10 and TC-12

Table 4
Classification accuracy of proposed and other methods for KTH-TIPS dataset

Table 5
Classification Accuracy of proposed and other methods for UIUC dataset