|
发布时间: 2018-07-16 |
图像分析和识别 |
|
|
收稿日期: 2017-07-17; 修回日期: 2018-01-25
基金项目: 国家自然科学基金项目(51275431);四川省科技支撑计划项目(2015GZ0200)
第一作者简介:
高攀, 1993年生, 男, 西安交通大学机械工程专业硕士研究生, 主要研究方向为图像识别、机器视觉。E-mail:nevermore_gp@163.com;
马子恒, 男, 硕士研究生, 研究方向为图像处理、模式识别。E-mail:zihengma@yeah.net; 于亚风, 男, 硕士研究生, 研究方向为机器视觉。E-mail:yuyafeng0917@163.com.
中图法分类号: TP391.5
文献标识码: A
文章编号: 1006-8961(2018)07-1024-09
|
摘要
目的 针对成对旋转不变的共生局部二值模式(PRICoLBP)算法对图像光照、旋转变化鲁棒性较差,且存在特征维度过大的问题,提出了一种可融合多种局部纹理结构信息的有效特征——增强成对旋转不变的共生扩展局部二值模式。方法 首先,对图像各像素点的邻域像素点灰度值进行二值量化得到二值编码序列,并不断旋转二值序列得到以不同邻域点作为编码起始点对应的LBP值;然后,分别利用极大、极小LBP值对应的邻域起始编码点和中心像素点确定两个方向矢量,并沿这两个方向矢量在两个不同尺度图像上选取上下文共生点;其次,利用扩展局部二值模式(ELBP)算法的旋转不变均匀描述子来提取上下文共生点对的中心像素灰度级、邻域像素灰度级及径向灰度差异特征间的相关性信息;最后,用上下文共生点对的特征直方图训练卡方核支持向量机,检测纹理图像类别。结果 通过对Brodatz、Outex(TC10、TC12)、Outex(TC14)、CUReT、KTH-TIPS和UIUC纹理库的分类实验,改进算法的识别率比原始的PRICoLBP算法识别率分别提高了0.32%、0.57%、5.62%、3.34%、2.1%、4.75%。结论 利用像素点LBP特征极值对应的起始编码序列来选取上下共生点对,并用ELBP算法提取共生点对局部纹理信息,故本文方法能更好描述共生点对间的高阶曲率信息及更多局部纹理信息。在具光照、旋转变化的Outex、CUReT、KTH-TIPS纹理库图像分类实验中,所提方法比原始PRICoLBP算法取得了更高的识别率。实验结果表明,改进算法相比于原始算法能在较低的特征维度下对图像光照、旋转变化具有较好的鲁棒性。
关键词
机器视觉; 模式识别; 局部二值模式; 空间上下文; 成对旋转不变; 极值; 鲁棒性
Abstract
Objective As effective local texture descriptors, local binary pattern (LBP) and its variants are widely applied in various fields of image processing. As an LBP variant, the pairwise rotation invariant co-occurrence LBP (PRICoLBP) algorithm classifies texture better than other LBP variant algorithms do by extracting high-order curvature and contextual co-occurrence information between spatial context co-occurrence pixel points. However, PRICoLBP determines these points by using the pixel gray gradient vector, which changes with image rotation and describes the image texture by using the co-occurrence histogram between the original LBP and uniform LBP (LBPu2) features of the context co-occurrence pixels. Consequently, the PRICoLBP algorithm has poor robustness to variations in image illumination and rotation and has high computing feature dimensionality. An efficient texture feature named enhanced pairwise rotation-invariant co-occurrence extended LBP (ELBP), which can fuse a variety of local texture structure information, is proposed in this work to address this problem. Method The binary coding sequence was obtained by performing binary quantization on the neighborhood pixel gray value of each pixel point. The LBP value corresponding to the different neighborhood points of each pixel, which was obtained by continuously rotating the binary coding sequence, was used as the initial point of coding. Then, two co-occurrence directional vectors with rotation invariance were determined using the central pixel point and the neighborhood initial points of coding corresponding to the maximum and minimum LBP values of each pixel, respectively. Two spatial context co-occurrence pixel points at different scales were selected along the two directional vectors on two grayscale images. Then, the correlation information among the central pixel gray level feature, the neighborhood pixel gray level feature, and the radial gray level difference feature of the spatial context co-occurrence points was extracted using the rotation-invariant uniform descriptor of the ELBP algorithm. The texture structure of a complex image was described by cascading the ELBP feature of each spatial context co-occurrence point. Finally, a chi-square kernel support vector machine, which was trained using texture feature histograms of the spatial context co-occurrence pixel pairs, was used to complete the detection of the image texture categories. Result Under the same experimental conditions, the classification recognition rate of the proposed method was improved by 0.32%, 0.57%, 5.62%, 3.34%, 2.1%, and 4.75% on the Brodatz, Outex (TC10, TC12), Outex (TC14), CUReT, KTH-TIPS, and UIUC texture databases, respectively, in comparison with the original PRICoLBP algorithm. Conclusion The initial encoding sequence corresponding to the maximum and minimum LBP feature values of each pixel point is used to select the spatial context co-occurrence pixel pairs, and the rotation-invariant uniform descriptor of the ELBP algorithm is adopted to capture the local texture structure information of the context co-occurrence pixel point pairs. Therefore, the proposed algorithm can describe the high-order curvature information and more local texture structure information between context co-occurrence pixel pairs better than the original PRICoLBP algorithm. In the image classification experiment on the Outex, CUReT, and KTH-TIPS image libraries, which have variations in texture illumination and rotation, the proposed algorithm not only exhibits a higher classification recognition rate than the original PRICoLBP algorithm but also has richer local texture feature patterns under reduced feature dimensionality. The experimental results show that the improved algorithm is more robust to variations in texture illumination and rotation than numerous state-of-the-art LBP variant algorithms under the same conditions. This algorithm can also be effectively applied in image classification with complex environment changes due to its high robustness and distinctivene.
Key words
machine vision; pattern recognition; local binary pattern (LBP); the spatial context; pairwise rotation invariant; extremum; robustness
0 引言
纹理作为图像重要的底层特征之一,是对物质表面固有属性的描述,在图像分析、机器视觉及模式识别领域占有重要地位,一直受到广大学者的关注与研究[1]。常用的纹理特征提取方法有灰度共生矩阵方法[2]、马尔可夫随机场方法[3]、小波变换方法[4]、分形理论[5]、基元字典的学习方法[6]等,但这些方法存在计算复杂度较大,对存在光照不均匀变化,背景干扰及噪声影响图像的分类效果欠佳等问题。
由于传统的纹理特征提取计算成本过高,影响到了纹理分类技术的实际应用,一些学者便开始简化图像纹理特征的计算方式。Ojala等人[7]提出了表征图像微观纹理的局部二值模式(LBP),因其原理简单、计算复杂度低而被广泛应用于图像分割、目标跟踪、人脸识别、医学图像分析等领域[8]。针对LBP算法对图像旋转、尺度、非均匀光照变化鲁棒性较差的问题,学者们提出了很多改进算法[9-16],比如,Tan等人[9]提出LTP算法引入
1 成对旋转不变的共生局部二值模式
1.1 局部二值模式
局部二值模式是一种典型的局部纹理描述子,通过比较图像任意像素点与其邻域像素点的灰度大小关系来编码图像局部纹理结构信息。如果邻域点灰度值大于等于中心像素点灰度值,则将该邻域像素点所在位置标记为1,否则置为0,按照规定的编码方向对不同的邻域点赋予不同的权重,将二进制序列转换一个无符号十进制数,并用该值作为像素点的LBP特征值。LBP特征的计算为
$ {L_{R, P}} = \sum\limits_{i = 0}^{P-1} {s\left( {{g_i}-{g_c}} \right){2^i}} $ | (1) |
$ s\left( x \right) = \left\{ \begin{gathered} 1\;\;x \geqslant 0 \hfill \\ 0\;\;x < 0 \hfill \\ \end{gathered} \right. $ | (2) |
式中,
图像旋转变化时会改变像素点LBP值,为了达到对图像的旋转不变性,Ojala等人[17]提出了具旋转不变性的LBP算子,即不断变换圆形邻域的起始编码序列得到一系列初始定义的LBP值,并将其中最小值作为该邻域的LBP值,其数学描述可表示为
$ L_{R, P}^{{\text{r}}i} = \min \left\{ {ROR\left( {{L_{R, P}}, i} \right)|i = 0, 1, \cdots, P-1} \right\} $ | (3) |
式中,
$ L_{R, P}^{{\text{riu}}2} = \left\{ \begin{gathered} \sum\limits_{i = 0}^{P-1} {s\left( {{g_i}-{g_c}} \right)} \;\;\;\;U\left( {{L_{R, P}}} \right) \leqslant 2 \hfill \\ P + 1\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;其他 \hfill \\ \end{gathered} \right. $ | (4) |
$ \begin{array}{l} U\left( {{L_{R, P}}} \right) = \left| {s\left( {{g_{p-1}}-{g_c}} \right)-s\left( {{g_0} - {g_c}} \right)} \right| + \\ \;\;\;\;\;\sum\limits_{i = 1}^{P - 1} {\left| {s\left( {{g_i} - {g_c}} \right) - s\left( {{g_{i - 1}} - {g_c}} \right)} \right|} \end{array} $ | (5) |
式中,
1.2 成对旋转不变的空间上下文共生特征
从信息论角度看,同时观察到两个具有一定相关性的事件发生比观察到它们独立发生能获取更多潜在有用信息。而应用到纹理分类上即认为:两种特征间的相关性特征能够描述比单个特征更大、更复杂的局部纹理结构。因此,文献[16]提出了一种有效的成对旋转不变的上下文纹理特征共生的编码策略,其思路为:对于图像中任意像素点
$ \left\{ \begin{array}{l} PROCoLB{P_{R, P}}\left( {A, B'} \right) = {\left[{L_{R, P}^{{\rm{u2}}}\left( A \right), L_{R, P}^{{\rm{riu2}}}\left( {B', i\left( A \right)} \right)} \right]_{co}}\\ i\left( A \right) = \mathop {\arg \;\max }\limits_{i \in \left\{ {0, 1, \cdots, P -1} \right\}} \left( {ROR\left( {{L_{R, P}}\left( A \right), i} \right)} \right)\\ B = a \cdot G\left( A \right) + b \cdot N\left( A \right) + A\\ B' = \phi \left( B \right) \end{array} \right. $ | (6) |
式中,
2 增强的成对旋转不变的共生扩展局部二值模式
由于原始PRICoLBP算法中共生点对位置选取对图像旋转变化较敏感,且上下文底层纹理特征的描述过于单一,对光照变化的鲁棒性较弱。针对这两个问题,设计了一种特征维度较低,对图像光照、旋转变化具有较强鲁棒性的算法,并将改进后的算法记为增强的成对旋转不变的共生扩展局部二值模式(EPRICoELBP)。
2.1 改进的空间上下文共生特征
由于图像旋转变化时,像素点周围的邻域像素会发生变化,即旋转变化后像素点的梯度方向也会变化,致使原始PRICoLBP算法对图像旋转变化的鲁棒性较差。而在不受噪声影响情况下,图像旋转变化可以视为仅改变了像素点LBP编码时邻域起始点位置,即像素LBP特征极大、极小值对应的邻域起始编码点并没有发生改变,如图 1所示。因此,分别以像素点和其LBP特征极值对应的邻域编码起始点来确定两个共生方向矢量
$ \left\{ \begin{array}{l} B = \phi \left( {A + a \cdot {f_{\max }}\left( A \right) + b \cdot {h_{\max }}\left( A \right)} \right)\\ C = \phi \left( {A + a \cdot {f_{\min }}\left( A \right) + b \cdot {h_{\min }}\left( A \right)} \right)\\ {i_{\max }}\left( A \right) = \mathop {\arg \;\max }\limits_{i \in \left\{ {0, 1, \cdots, P-1} \right\}} \left\{ {ROR\left( {{L_{R, P}}\left( A \right), i} \right)} \right\}\\ {i_{\min }}\left( A \right) = \mathop {\arg \;\min }\limits_{i \in \left\{ {0, 1, \cdots, P-1} \right\}} \left\{ {ROR\left( {{L_{R, P}}\left( A \right), i} \right)} \right\} \end{array} \right. $ | (7) |
式中,
2.2 融合ELBP的局部纹理特征
文献[15]为解决LBP算法对纹理非线性光照和旋转变化敏感的问题,提出了一种扩展的局部二值模式(ELBP),其利用两种具有互补性的像素灰度和差分特征的联合分布形式来表示图像的局部纹理结构。其中,像素灰度特征包括基于邻域点灰度的局部二值模式(NI-LBP)、中心像素点灰度特征(CI-LBP);像素灰度差分特征包括径向灰度差分局部二值模式(RD-LBP)和角向灰度差分局部二值模式(AD-LBP)。本文引进ELBP算法,分别提取像素点的局部邻域点灰度值特征、径向邻域点灰度值差异特征和中心像素点灰度特征,相较于原始算法,改进算法拥有更丰富的特征模式。ELBP数学描述可表示为
$ \left\{ \begin{array}{l} E = N/R/C\\ {N_{R, P}} = \sum\limits_{i = 0}^{P-1} {s\left( {{g_i}-{\mu _P}} \right){2^i}, {\mu _P} = \frac{1}{P}\sum\limits_{i = 0}^{P-1} {{g_i}} } \\ {R_{R, P}} = \sum\limits_{i = 0}^{P - 1} {s\left( {\Delta _{\delta, {\rm{i}}}^{{\rm{Rad}}}} \right){2^i}, \Delta _{\delta, {\rm{i}}}^{{\rm{Rad}}} = {x_{R, i}} - {x_{R - \delta, i}}} \\ {C_{R, P}} = s\left( {{g_c} - {\mu _I}} \right), {\mu _I} = mean\left( {I\left( : \right)} \right) \end{array} \right. $ | (8) |
式中,
为提升算法计算效率,采用对光照、旋转变化鲁棒性较强的旋转不变均匀模式来对ELBP特征进行归类[18],即利用式(4)(5)将上下文共生像素点的局部邻域像素灰度特征和径向灰度差分特征各归为10类模式,并将归类后得到的
$ \left\{ \begin{array}{l} {\rm{EPRICoELBP}}\left( A \right) = {\left[{\varphi \left( A \right), \varphi \left( {B'} \right), \varphi \left( {C'} \right)} \right]_{co}}\\ \varphi \left( A \right) = {E^{{\rm{riu2}}}}\left( A \right)\\ \varphi \left( {B'} \right) = {E^{{\rm{riu2}}}}\left( {B', {i_{\max }}\left( A \right)} \right)\\ \varphi \left( {C'} \right) = {E^{{\rm{riu2}}}}\left( {C', {i_{\max }}\left( A \right)} \right) \end{array} \right. $ | (9) |
2.3 本文算法计算流程
1) 初始化共生特征直方图
2) 将原始图像
3) 根据式(7)分别计算出
4) 根据式(8)分别计算出3个像素点的
$ \left\{ \begin{array}{l} x = {N^{{\rm{riu2}}}}\left( Q \right)\\ y = {R^{{\rm{riu}}2}}\left( Q \right)\\ z = C\left( Q \right)\\ H\left( {x, y, z, n} \right) = H\left( {x, y, z, n} \right) + \\ M\left( Q \right), n = 1, 2, 3 \end{array} \right. $ | (10) |
式中,
5) 遍历数据库中纹理图像,将得到的特征直方图送入支持向量机(SVM)进行模板训练与纹理分类识别。利用改进算法进行图像纹理分类的实验过程如图 2所示。
3 实验结果与分析
3.1 实验参数设置及数据选取
为验证本文提出算法对光照、旋转变化的鲁棒性,选择在标准数据库Brodatz、Outex、CUReT、KTH-TIPS及UIUC上分别与LBP、LBPV[19]、CLBP、LTP、CLBC、PRICoLBP、ELBP算法进行对比。为保证对比试验公平性,LBP及其改进算法均采用旋转不变均匀模式,各算法的采样半径和邻域点数分别为2和8,LTP算法的阈值波动区间为5,NRLBP方法的区间值设置为3,实验基于Vlfeat0.9.20开源函数库,采用卡方核支持向量机[20]评估各算法的分类性能。为避免随机偏差,在各数据库上进行纹理分类实验时,采用100次独立的随机采样实验的平均分类正确率作为评估算法性能的指标。
实验所用的各数据库样本纹理特征如表 1所示,其中Brodatz数据库纹理均拍摄于同一光照、角度、尺度下,可用于评估算法区分性;TC10、TC12和UIUC数据库纹理都有明显旋转变化,主要测试算法对旋转变化的鲁棒性;TC14库的各类图像均拍摄于不同的光照条件,主要用于评估算法对光照变化的鲁棒性;而CUReT、KTH-TIPS数据库纹理包含了复杂的光照、旋转变化,主要用于测试算法对复杂环境变化的鲁棒性。Outex、CUReT数据库中部分纹理样本如图 3、图 4所示。
表 1
实验使用的各纹理库特征
Table 1
Characteristics of the texture datasets used in our experiments
纹理图像库 | 光照变化 | 旋转变化 | 尺度变化 | 图像数目 | 纹理类别数 | 训练样本数目 |
Brodatz纹理库 | × | × | × | 896 | 112 | 4 |
Outex纹理库(TC10) | × | √ | × | 4 320 | 24 | 20 |
Outex纹理库(TC12) | √ | √ | × | 4 800 | 24 | 20 |
Outex纹理库(TC14) | √ | × | × | 4 080 | 68 | 30 |
KTH_TIPS纹理库 | √ | √ | √ | 810 | 10 | 40 |
CUReT纹理库 | √ | √ | × | 5 612 | 61 | 46 |
UIUC纹理库 | × | √ | √ | 1 000 | 24 | 20 |
3.2 实验结果分析与讨论
从表 2和表 3中的实验数据来看,本文EPRICoELBP算法在Brodatz、OuTex、UIUC、KTH-TIPS纹理库上识别率均高于其余LBP算法,说明EPRICoELBP算法具有较强的特征鉴别能力;而在具纹理光照多样性和旋转变化的OuTex、UIUC和KTH-TIPS数据库上,改进后的EPRICoELBP算法识别率明显高于其余算法,说明EPRICoELBP算法对光照、旋转变化的鲁棒性较强。
表 2
不同算法在Brodatz、TC14、UIUC、KTH-TIPS纹理库中识别率
Table 2
The recognition rates of different algorithm on Brodatz、TC14、UIUC、KTH-TIPS database
/% | ||||||||
数据库 | LBP | LTP | LBPV | CLBP | CLBC | PRICoLBP | ELBP | EPRICoELBP |
Brodatz | 78.16 | 96.81 | 87.38 | 96.66 | 95.71 | 97.38 | 97.21 | 97.69 |
TC14 | 67.06 | 83.98 | 71.65 | 91.82 | 91.31 | 88.14 | 90.34 | 93.10 |
UIUC | 78.24 | 92.35 | 82.61 | 96.74 | 96.41 | 92.90 | 96.42 | 97.32 |
KTH-TIPS | 76.95 | 94.46 | 94.12 | 97.05 | 97.48 | 95.83 | 97.10 | 97.91 |
表 3
不同算法在OuTex(TC10、TC12)纹理库上的识别率
Table 3
The recognition rates of different algorithm on OuTex(TC10、TC12) database
/% | ||||
算法 | TC10 | TC12 | 平均 | |
horizon | Tl84 | |||
LBP | 87.39 | 84.03 | 88.62 | 86.68 |
LBPV | 97.62 | 93.65 | 95.84 | 95.70 |
LTP | 99.18 | 97.82 | 98.43 | 98.48 |
CLBP | 99.36 | 98.68 | 99.32 | 99.12 |
CLBC | 99.65 | 99.11 | 99.45 | 99.40 |
ELBP | 99.71 | 99.33 | 99.67 | 99.57 |
PRICoLBP | 99.40 | 98.85 | 99.28 | 99.18 |
EPRICoELBP | 99.85 | 99.62 | 99.78 | 99.75 |
从表 2分类数据可以看出,对于样本均拍摄于相同光照、视角和尺度条件下的Brodatz数据库,本文EPRICoELBP算法的识别率相比于原始PRICoLBP算法识别率提升了0.32%,比ELBP算法提升了0.49%,说明通过融合上下文共生像素点的局部ELBP特征,可有效增强算法的区分性能;对于仅存在光照条件变化的TC14纹理库分类,本文提出的EPRICoELBP算法分类识别率比CLBP、CLBC、PRICoLBP、ELBP算法分别提高了1.39%、1.96%、5.62%、2.96%,说明EPRICoELBP算法对光照变化的鲁棒性较强;对于存在尺度、视角变化的UIUC纹理库,EPRICoELBP算法识别率比CLBP、CLBC、PRICoLBP、ELBP算法分别高出了0.6%、0.94%、4.75%、0.93%,说明本文EPRICoELBP算法对纹理旋转变化具有较强的鲁棒性;特别是在样本同时存在光照、旋转和尺度变化的KTH-TIPS数据库上,EPRICoELBP算法识别率比CLBP、CLBC、PRICoLBP、ELBP各高出了0.89%、0.44%、2.1%、0.83%,说明EPRICoELBP算法利用像素局部特征极值对应的邻域起始编码点与邻域中心像素点来确定上下文共生点对,并采用ELBP算法提取上下文共生点的邻域灰度信息和径向灰度差值信息,可有效增强算法对光照、旋转变化的鲁棒性。
从表 3分类数据可以看出,对于存在明显旋转变化的TC10和TC12纹理库分类,原始PRICoLBP算法的平均识别率都高于LBP及其改进算法,说明通过编码共生点对间的高阶夹角信息可提升算法对旋转变化的鲁棒性。但原始PRICoLBP算法计算上下文共生点对位置的梯度矢量会随图像旋转变化而变化,致使提取到的高阶夹角信息会因图像旋转变化而失真,而本文EPRICoELBP算法采用极值化像素点LBP特征值策略来确定上下文共生矢量方向,确保了算法提取到的上下文共生点对间的高阶夹角信息的稳定性,使得改进后的EPRICoELBP算法相比原始PRICoLBP算法平均识别率高出了0.57%,有效说明了本文EPRICoELBP算法比原始PRICoLBP算法对图像旋转变化具有更高的鲁棒性。
表 4主要验证了本文算法对外界复杂环境变化的鲁棒性。从表 4分类数据可以看出,样本光照、旋转变化的多样性并没影响到本文算法的分类性能,在不同训练样本数量的条件下,本文EPRICoELBP算法相较于各类LBP改进算法均取得了较好分类效果,其最高识别率比CLBP、CLBC、PRICoLBP、ELBP算法识别率分别提高了1.28%、1.45%、3.34%、1.72%,说明EPRICoELBP算法对复杂外界环境变化的鲁棒性较高。
表 4
不同算法在CUReT数据库上的识别率
Table 4
The recognition rates of different algorithm on CUReT database
/% | ||||
算法 | 训练样本数 | |||
N=6 | N=12 | N=23 | N=46 | |
LBP | 41.20 | 46.78 | 52.64 | 58.09 |
LTP | 72.57 | 83.34 | 89.52 | 93.23 |
LBPV | 67.88 | 78.37 | 85.08 | 89.48 |
CLBP | 82.40 | 91.16 | 95.57 | 97.68 |
CLBC | 82.21 | 90.62 | 95.06 | 97.39 |
ELBP | 84.23 | 91.24 | 95.04 | 97.13 |
PRICoLBP | 78.50 | 87.02 | 92.23 | 95.60 |
EPRICoELBP | 84.60 | 92.38 | 96.39 | 98.80 |
图 5给出了各类LBP改进算法在不同数据库上的纹理分类识别率对比结果。从中可以看出,在样本具有光照、旋转变化的Outex、UIUC、CUReT、KTH-TIPS数据库上,EPRICoELBP算法识别率均高于其余LBP算法,说明EPRICoELBP算法对图像光照、旋转变化的鲁棒性较强;同时在具有复杂纹理变化的CUReT和KTH-TIPS数据库上,EPRICoELBP算法识别率均高于PRICoLBP算法和ELBP算法,表明通过计算像素局部特征极值对应的邻域起始编码点来选取不同尺度上的上下文共生点对及融合具有互补性的共生点对的ELBP纹理特征,有效提升了原始PRICoLBP算法对光照、旋转变化的鲁棒性。
表 5给出了不同算法在KTH-TIPS数据库上的计算复杂度对比情况,其中分类时间表示进行100次特征分类实验的总时间均值,包括了分类器(SVM)模型训练和特征标签预测的时间。从表中数据可以看出,原始PRICoLBP算法和EPRICoELBP算法虽然特征维度比其余LBP算法特征维度高,但由于其利用上下文共生点对特征来描述纹理结构,具有丰富的特征模式,使得这两种算法的特征分类时间都较少;而本文EPRICoELBP算法采用ELBP算子提取上下文共生点的邻域灰度信息和径向灰度差分信息,并利用旋转不变均匀模式来编码单个
表 5
不同算法在KTH-TIPS数据库上计算复杂度
Table 5
The computational complexity of different algorithm on KTH-TIPS database
算法 | 特征维度 | 分类时间/(ms/次) |
LBP | 10 | 262.75 |
LTP | 100 | 60.56 |
LBPV | 10 | 479.72 |
CLBP | 200 | 42.69 |
CLBC | 162 | 44.78 |
ELBP | 200 | 43.38 |
PRICoLBP | 1 180 | 63.59 |
EPRICoELBP | 600 | 47.92 |
4 结论
本文针对原始PRICoLBP算法对图像旋转、光照变化不稳定及特征维度过大等问题,提出了改进的EPRICoELBP算法。首先通过极值化像素LBP特征值来计算其对应邻域起始编码点坐标,并利用中心像素点坐标和极值化LBP特征得到的邻域起始编码点坐标来确定两个上下文共生的方向矢量;其次在这两个方向矢量上选取不同尺度的共生点对,并将其映射到对应的尺度空间图像上;最后利用ELBP算法提取到的3个上下文共生点的纹理特征的联合分布形式描述图像局部纹理结构。在Brodatz、OuTex、CUReT、KTH-TIPS、UIUC纹理库上的分类实验表明,本文算法比原始的PRICoLBP算法及各种LBP改进算法取得了更好的分类性能,说明改进后的EPRICoELBP算法通过提取更多的空间上下文纹理特征及高阶曲率信息,有效增强了算法对光照、旋转变化的鲁棒性,同时也大大降低了算法的计算复杂度。由于EPRICoELBP算法性能评估均在公开的标准纹理库上进行,其中的样本均不包含噪声,而实际生活中拍摄的纹理样本均包含不同性质、不同强度的噪声,缺乏对算法抗噪性能的讨论,因此后续工作主要将围绕算法对各种噪声的抗噪性能分析而展开。
参考文献
-
[1] Liu L, Kuang G Y. Overview of image textural feature extraction methods[J]. Journal of Image and Graphics, 2009, 14(4): 622–635. [刘丽, 匡纲要. 图像纹理特征提取方法综述[J]. 中国图象图形学报, 2009, 14(4): 622–635. ] [DOI:10.11834/jig.20090409]
-
[2] Vargas J F, Ferrer M A, Travieso C M, et al. Off-Line signature verification based on grey level information using texture features[J]. Pattern Recognition, 2011, 44(2): 375–385. [DOI:10.1016/j.patcog.2010.07.028]
-
[3] Huang A, Abugharbieh R, Tam R. A novel rotationally invariant region-based hidden Markov model for efficient 3-D image segmentation[J]. IEEE Transactions on Image Processing, 2010, 19(10): 2737–2748. [DOI:10.1109/TIP.2010.2048965]
-
[4] Choy S K, Tong C S. Statistical wavelet subband characterization based on generalized gamma density and its application in texture retrieval[J]. IEEE Transactions on Image Processing, 2010, 19(2): 281–289. [DOI:10.1109/TIP.2009.2033400]
-
[5] Galloway M M. Texture analysis using gray level run lengths[J]. Computer Graphics and Image Processing, 1975, 4(2): 172–179. [DOI:10.1016/S0146-664X(75)80008-6]
-
[6] Varma M, Zisserman A. A statistical approach to texture classification from single images[J]. International Journal of Computer Vision, 2005, 62(1-2): 61–81. [DOI:10.1023/B:VISI.0000046589.39864.ee]
-
[7] Ojala T, Pietikäinen M, Harwood D. A comparative study of texture measures with classification based on featured distributions[J]. Pattern Recognition, 1996, 29(1): 51–59. [DOI:10.1016/0031-3203(95)00067-4]
-
[8] Song K C, Yan Y H, Chen W H, et al. Research and Perspective on local binary pattern[J]. Acta Automatica Sinica, 2013, 39(6): 730–744. [宋克臣, 颜云辉, 陈文辉, 等. 局部二值模式方法研究与展望[J]. 自动化学报, 2013, 39(6): 730–744. ] [DOI:10.3724/SP.J.1004.2013.00730]
-
[9] Tan X Y, Triggs B. Enhanced local texture feature sets for face recognition under difficult lighting conditions[J]. IEEE Transactions on Image Processing, 2010, 19(6): 1635–1650. [DOI:10.1109/TIP.2010.2042645]
-
[10] Ren J F, Jiang X D, Yuan J S. Noise-resistant local binary pattern with an embedded error-correction mechanism[J]. IEEE Transactions on Image Processing, 2013, 22(10): 4049–4060. [DOI:10.1109/TIP.2013.2268976]
-
[11] Guo Z H, Zhang L, Zhang D. A completed modeling of local binary pattern operator for texture classification[J]. IEEE Transactions on Image Processing, 2010, 19(6): 1657–1663. [DOI:10.1109/TIP.2010.2044957]
-
[12] Zhao Y, Huang D S, Jia W. Completed local binary count for rotation invariant texture classification[J]. IEEE Transactions on Image Processing, 2012, 21(10): 4492–4497. [DOI:10.1109/TIP.2012.2204271]
-
[13] Heikkilä M, Pietikäinen M, Schmid C. Description of interest regions with local binary patterns[J]. Pattern Recognition, 2009, 42(3): 425–436. [DOI:10.1016/j.patcog.2008.08.014]
-
[14] Liao S, Law M W K, Chung A C S. Dominant local binary patterns for texture classification[J]. IEEE Transactions on Image Processing, 2009, 18(5): 1107–1118. [DOI:10.1109/TIP.2009.2015682]
-
[15] Liu L, Zhao L J, Long Y L, et al. Extended local binary patterns for texture classification[J]. Image and Vision Computing, 2012, 30(2): 86–99. [DOI:10.1016/j.imavis.2012.01.001]
-
[16] Qi X B, Xiao R, Li C G, et al. Pairwise rotation invariant co-occurrence local binary pattern[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2014, 36(11): 2199–2213. [DOI:10.1109/TPAMI.2014.2316826]
-
[17] Ojala T, Pietikäinen M, Maenpaa T. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2002, 24(7): 971–987. [DOI:10.1109/TPAMI.2002.1017623]
-
[18] Liu L, Xie Y X, Wei Y M, et al. Survey of local binary pattern method[J]. Journal of Image and Graphics, 2014, 19(12): 1696–1720. [刘丽, 谢毓湘, 魏迎梅, 等. 局部二进制模式方法综述[J]. 中国图象图形学报, 2014, 19(12): 1696–1720. ] [DOI:10.11834/jig.20141202]
-
[19] Guo Z H, Zhang L, Zhang D. Rotation invariant texture classification using LBP variance (LBPV) with global matching[J]. Pattern Recognition, 2010, 43(3): 706–719. [DOI:10.1016/j.patcog.2009.08.017]
-
[20] Akata Z, Perronnin F, Harchaoui Z, et al. Good practice in large-scale learning for image classification[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2014, 36(3): 507–520. [DOI:10.1109/TPAMI.2013.146]