﻿ REM记忆模型在图像分类识别中的应用
«上一篇
 文章快速检索 高级检索

 智能系统学报  2017, Vol. 12 Issue (3): 310-317  DOI: 10.11992/tis.201605010 0

### 引用本文

JIANG Ying, WANG Yanjiang. Application of REM memory model in image recognition and classification[J]. CAAI Transactions on Intelligent Systems, 2017, 12(3): 310-317. DOI: 10.11992/tis.201605010.

### 文章历史

REM记忆模型在图像分类识别中的应用

Application of REM memory model in image recognition and classification
JIANG Ying, WANG Yanjiang
College of Information and Control Engineering, China University of Petroleum, Qingdao 266580, China
Abstract: We attempt to combine a memory model with image learning and recognition and to research the application of the REM model in image recognition and classification. An image feature vector was obtained by histograms of oriented gradients (HOG) and local binary pattern (LBP) operators; every component of a feature vector was copied with a certain probability, allowing for an error-prone copy of the studied vector. Finally, Bayesian decision theory was applied for calculating the average likelihood ratio between the feature vector of the probe image and that of the studied image set. The value of the ratio was used to decide whether the probe image had been studied.Experimental results demonstrate that the proposed method can gain a good recognition effect not only for the classification of the same object with small rotation angles but also for the recognition of the same category object. Moreover, the false rate is far lower than that of other classification methods.
Key words: image recognition    memory modeling    HOG feature    LBP feature    Bayesian decision

1 图像特征表达

1.1 HOG特征

 图 1 HOG特征提取算法流程 Fig.1 The flow chart of HOG algorithm
1.2 LBP特征

1) 对图像中的每一个像素点，定义圆形邻域窗口，每个像素的灰度值与其相邻的8个像素的灰度值比较，若周围像素值大于中心像素值，则该像素点的位置被标记为1，否则为0。这样可产生8位二进制数，即得到该窗口中心像素点的初始LBP值。

2) 不断旋转圆形邻域得到一系列初始定义的LBP值，取最小值作为该像素点的LBP值。

3) 统计LBP值对应的二进制数从0~1或1~0跳变的次数，根据跳变次数确定其属于哪一种LBP模式，共有P+1=9种模式，得到的模式数值即为像素点的LBP值。

4) 图像中所有像素点的LBP值组合起来形成一个LBP特征矩阵，即为该图像的LBP特征。

2 REM模型在视觉图像的表达、存储与提取中的应用

REM记忆模型被提出之后，研究人员陆续对REM模型进行研究。Stams等[18]通过对编码与提取过程中的项目强度的控制，对比研究了REM模型与BCDMEM模型对提取过程中项目强度对误报率降低的解释说明。Cox等[19]在REM与RCA-REM模型基础上提出一个新认知记忆模型，证实了即使在任务、学习因素、刺激及其他因子变化情况下，所提方法都有可能获得合理的认知决策。Criss等[20]对比了REM模型与SLiM(the subjective likelihood model)模型，发现REM模型预测的误报率较高；M.Montenegro等[21]研究了REM模型的解析表达式，文中引入Fourier变换，给出REM模型的FT积分方程，导出在给定参数值下模型预测的命中率与误报率的双积分形式的解析表达式，同时发现其具有与BCDMEM模型相同的一些性质：模型是不确定的，除非其中的一个参数固定为一个预设值，向量长度参数是不可忽略的参数。

2.1 特征表达与存储

REM模型指出人脑记忆由图像构成，每幅图像是由一个特征值向量表示的，并且最终存储结果是对特征值向量的一个不完整且容易出错的复制。本文试图借鉴REM模型对单词的存储学习过程来模拟人脑对图像的学习过程，有概率地对图像的特征向量进行复制，同时在复制过程中允许出现错误值。

V={Vj}j=1, 2, …, N标记所有已学习图像的特征集，其中Vj表示已学习图像集合中第j副图像Ij的特征向量，N为已学习图像集合中的图像个数。

2.2 提取

 $\begin{array}{l} \;\;\;\;\;\;\;\;\;\;\;{\lambda _j} = \frac{{P\left( {{D_j}\left| {{S_j}} \right.} \right)}}{{P\left( {{D_j}\left| {{N_j}} \right.} \right)}} = \\ {\left( {1-c} \right)^{{n_{jq}}}}\prod\limits_{k \in M} {\frac{{c + \left( {1-c} \right)g{{\left( {1-g} \right)}^{{V_{kj}} - 1}}}}{{g{{\left( {1 - g} \right)}^{{V_{kj}} - 1}}}}} \end{array}$ (1)

2.3 Bayesian决策

 $\begin{array}{l} \phi = \frac{{P\left( {O\left| D \right.} \right)}}{{P\left( {N\left| D \right.} \right)}} = \frac{{\frac{{P\left( O \right)P\left( {D\left| O \right.} \right)}}{{P\left( D \right)}}}}{{\frac{{P\left( N \right)P\left( {D\left| N \right.} \right)}}{{P\left( D \right)}}}} = \frac{{P\left( {D\left| O \right.} \right)}}{{P\left( {D\left| N \right.} \right)}} = \\ \;\;\;\;\;\;\sum\limits_{j = 1}^N {\frac{{P\left( {D\left| {{S_j}} \right.} \right)P\left( {{S_j}} \right)}}{{P\left( {D\left| N \right.} \right)}} = } \sum\limits_{j = 1}^N {\frac{1}{N}\frac{{P\left( {D\left| {{S_j}} \right.} \right)}}{{P\left( {D\left| N \right.} \right)}} = } \\ \;\;\;\;\;\;\frac{1}{N}\sum\limits_{j = 1}^N {\frac{{P\left( {{D_j}\left| {{S_j}} \right.} \right)\prod\limits_{i \ne j} {P\left( {{D_i}\left| {{N_i}} \right.} \right)} }}{{P\left( {{D_j}\left| {{N_j}} \right.} \right)\prod\limits_{i \ne j} {P\left( {{D_i}\left| {{N_i}} \right.} \right)} }} = } \\ \;\;\;\;\;\;\frac{1}{N}\sum\limits_{j = 1}^N {\frac{{P\left( {{D_j}\left| {{S_j}} \right.} \right)}}{{P\left( {{D_j}\left| {{N_j}} \right.} \right)}} = } \frac{1}{N}{\lambda _j} \end{array}$

ϕ>1，那么认为被检测的图像为已学习的图像，同时认为该图像匹配最大λj值对应第j幅图像；反之认为被检测图像是新的，从未学习过。

 图 2 改进的图像特征存储提取数值例子 Fig.2 An improved numerical example for the storage and retrieval of the image feature
3 实验结果

3.1 Coil-20数据库实验结果

Coil-20数据库由20个不同对象的旋转图像构成，每个对象在水平方向上旋转360°，并每隔5°拍摄一张照片，故每个项目有72幅图像，每幅图像的像素为128×128。

 图 3 已学习图像集 Fig.3 The studied images set

 图 4 被测试图像集 Fig.4 The probe images set

3.2 Caltech-256数据库实验结果

Caltech-256数据库来自加利福利亚理工学院，该数据库共有29 780副图像，包含了256个不同图像项目类别，每个图像类别包含不少于80幅属于该类别的不同图像，这些图像属于同一类，但并不是完全相同的项目，实验选择的已学习图像列表如图 5所示。

 图 5 Caltech-256数据库中已学习图像集 Fig.5 The studied images set on the Caltech-256 database

 图 6 Caltech-256数据库中被检测图像集 Fig.6 The probe images set on the Caltech-256 database

3.2.1 实验参数对识别率的影响

 图 7 不同Φ值与LBP倍数变化下的实验曲线 Fig.7 The hit and false rate curve under the varying value of Φ and LBP

3.2.2 与SRC算法的实验性能对比

 $\mathop {\min }\limits_H \left\| {\mathit{\boldsymbol{Y}}-\mathit{\boldsymbol{X}}H} \right\|_2^2$ (3)

3.2.3 与支持向量机(SVM)算法的实验性能对比

4 结束语

 [1] XU Y, ZHANG B, ZHONG Z. Multiple representations and sparse representation for image classification[J]. Pattern recognition letters, 2015, 68: 9-14. DOI:10.1016/j.patrec.2015.07.032 (0) [2] HELLMAN M E. The nearest neighbor classification rule with a reject option[J]. IEEE transactions on systems science and cybernetics, 1970, 6(3): 179-185. DOI:10.1109/TSSC.1970.300339 (0) [3] DENOEUX T. A neural network classifier based on Dempster-Shafer theory[J]. IEEE transactions on systems man and cybernetics-Part A systems and humans, 2010, 30(2): 131-150. (0) [4] DAVTALAB R, DEZFOULIAN M H, MANSOORIZADEH M. Multi-level fuzzy min-max neural network classifier[J]. IEEE transactions on neural networks and learning systems, 2014, 25(3): 470-482. DOI:10.1109/TNNLS.2013.2275937 (0) [5] CHANG C C, LIN C J. LIBSVM: a library for support vector machines[J]. ACM transactions on intelligent systems and technology, 2011, 2(3): 389-396. (0) [6] SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[J]. Computer science, 2014. (0) [7] HUANG G B. What are extreme learning machines? Filling the gap between Frank Rosenblatt's dream and John von Neumann's puzzle[J]. Cognitive computation, 2015, 7(3): 263-278. DOI:10.1007/s12559-015-9333-0 (0) [8] WRIGHT J, YANG A Y, GANESH A, et al. Robust face recognition via sparse representation[J]. IEEE transactions on pattern analysis and machine intelligence, 2008, 31(2): 210-227. (0) [9] ZHANG L, YANG M, FENG X. Sparse representation or collaborative representation: Which helps face recognition[C]//International Conference on Computer Vision, 2011: 471-478. http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6126277 (0) [10] YANG M, ZHANG L YANG J, ZHANG D. Robust sparse coding for face recognition[C]//IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), 2011, 42(7): 625-632. http://ieeexplore.ieee.org/xpls/icp.jsp?arnumber=5995393 (0) [11] SCHOLKOPF B, PLATT J, HOFMANN T. Sparse representation for signal classification[J]. Advances in neural information processing systems, 2007, 19: 609-616. (0) [12] WANG Y. Formal description of the cognitive process of memorization[M]//Transactions on Computational Science V. Springer Berlin Heidelberg, 2009: 81-98. (0) [13] WANG Y, CHIEW V. On the cognitive process of human problem solving[J]. Cognitive systems research, 2010, 11(1): 81-92. DOI:10.1016/j.cogsys.2008.08.003 (0) [14] MURDOCK B B. A theory of the storage and retrieval of item and associative information[J]. Psychological review, 1982, 89(6): 609-626. DOI:10.1037/0033-295X.89.6.609 (0) [15] RAAIJMAKERS J G W, SHIFFRIN R M. SAM: a theory of probabilistic search of associative memory[J]. Psychology of learning and motivation, 1980, 14: 207-262. DOI:10.1016/S0079-7421(08)60162-0 (0) [16] HINTZMAN D L. Judgments of frequency and recognition memory in a multiple-trace model[J]. Psychological review, 1988, 95(4): 528-551. DOI:10.1037/0033-295X.95.4.528 (0) [17] SHIFFRIN R M, STEYVERS M. A model for recognition memory: REM-retrieving effectively from memory[J]. Psychonomic bulletin and review, 1997, 4(2): 145-166. DOI:10.3758/BF03209391 (0) [18] STARN J J, WHITE C N, RATCLIFF R. A direct test of the differentiation mechanism: REM, BCDMEM, and the strength-based mirror effect in recognition memory[J]. Journal of memory and language, 2010, 63(1): 18-34. DOI:10.1016/j.jml.2010.03.004 (0) [19] COX G E, SHIFFRIN R M. Criterion Setting and the dynamics of recognition memory[J]. Topics in cognitive science, 2012, 4(1): 135-150. DOI:10.1111/tops.2012.4.issue-1 (0) [20] CRISS A H, MCCLELLAND J L. Differentiating the differentiation models: a comparison of the retrieving effectively from memory model (REM) and the subjective likelihood model (SLiM)[J]. Journal of memory and language, 2006, 55(4): 447-460. DOI:10.1016/j.jml.2006.06.003 (0) [21] MONTENEGRO M, MYUNG J I, PITT M A. Analytical expressions for the REM model of recognition memory[J]. Journal of mathematical psychology, 2014, 60(3): 23-28. (0) [22] DALAL N, TRIGGS B. Histograms of oriented gradients for human detection[C]//Computer Vision and Pattern Recognition. IEEE Computer Society Conference, on, 2005: 886-893. http://doi.ieeecomputersociety.org/resolve?ref_id=doi:10.1109/CVPR.2005.177&rfr_id=trans/tp/2008/10/ttp2008101713.htm (0) [23] OJALA T, HARWOOD I. A Comparative study of texture measures with classification based on feature distributions[J]. Pattern recognition, 1996, 29(1): 51-59. DOI:10.1016/0031-3203(95)00067-4 (0) [24] LOWE D G. Distinctive image features from scale-invariant keypoints[J]. International journal of computer vision, 2004, 60(2): 91-110. DOI:10.1023/B:VISI.0000029664.99615.94 (0) [25] BAY H, TUYTELAARS T, VAN GOOL L. SURF: speeded up robust features[J]. Computer vision and image understanding, 2006, 110(3): 404-417. (0) [26] PIETIKAINEN M, OJALA T, XU Z. Rotation-invariant texture classification using feature distributions[J]. Pattern recognition, 2000, 33(1): 43-52. DOI:10.1016/S0031-3203(99)00032-1 (0) [27] OJALA T, PIETIKAINEN, 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 (0) [28] NENE A S, NAYAR S K, MURASE H. Columbia object image library (COIL-20)[R]. CUCS-005-96, 1996. http://www.researchgate.net/publication/2784735_Columbia_Object_Image_Library_ (0) [29] GRIFFIN G, HOLUB A, PERONA P. Caltech-256 object category dataset[R]. Pasadena: California Institute of Technology, 2007. https://www.researchgate.net/publication/30766223_Caltech-256_Object_Category_Dataset (0)