Multimedia Devices Laboratory

Qiangfu Zhao Professor	Yong Liu Associate Professor	Lee Yang-Han Visiting Researcher

The main stream in our lab is related to computational intelligence. So far we have focused our study on three key words: recognition, learning and understanding. The goal of our research is to develop some learning models that are flexible enough to adapt changing environment, and also simple enough to be realized, interpreted and re-used. The ultimate goal is to design a system that can think, and decide what to do and how to grow-up based on its own thinking. For this purpose, many approaches have been studied - e.g., neuro-computation, evolutionary computation, reinforcement learning, and so on. Of course, results proposed in conventional symbol based artificial intelligence are also included. So far we have used or proposed the following learning models: Neural network trees (NNTrees), Nearest neighbor classifier trees (NNC-Trees), Neural network ensembles, Modular neural networks, Cellular automata, and Recurrent neural networks. Based on the above learning models, we have proposed many new algorithms. Examples include: IEA: individual evolutionary algorithm (also called the R4-rule), CoopCEA: cooperative co-evolutionary algorithms, EPNet: evolutionary programming neural net, and Evolutionary design of neural network trees. Induction of NNC-Trees with the R4-rule. Fast neural network for face detection, and Rough null space appraoch to face recognition. To verify and to improve the models and learning algorithms proposed so far, we have being studying on-line growing of neural network trees, evolution of neural network ensemble, evolutionary design of decision trees, and so on. Currently, we are very interested in applying our models and algorithms to solving practical problems related to producing a "safe, secure and healthy" society. Examples include: face detection, face/expression recognition, image compression and protection, medical diagnosis, and text mining.

Refereed Journal Papers

[qf-zhao-01:2007]

Q. F. Zhao. k-PCA: a semi-universal encoder for image compression. International Journal of Pervasive Computing and Communications, 3(2):205-220, 2007.

In recent years, principal component analysis (PCA) hasattracted great attention in image compression. However, since the compressedimage data include both the transformation matrix (the eigenvectors)and the transformed coefficients, PCA cannot produce highcompression ratio. To solve this problem, we propose k-PCA that is acombination of vector quantization (VQ) and PCA. The basic idea isto divide the problem space into k clusters using VQ, and then finda set of eigenvectors using PCA for each cluster. The point is thatif the k-PCA is obtained using data containing enough information,the k-PCA can be used as a universal encoder to compress any imagein a given domain. Thus, although k-PCA is more complex than asingle PCA, the compression ratio can be much higher because theeigenvectors can be excluded from the encoded data. The performanceof the k-PCA can be improved further through learning. For thispurpose, an extended LBG algorithm is proposed in this paper. Theeffectiveness of the k-PCA is shown through experiments with severalwell-known test images.

Refereed Proceeding Papers

[qf-zhao-02:2007]	H. Kobayashi and Q. F. Zhao. Face Detection with Clustering, LDA and NN. In IEEE, editor, Proc. IEEE International Conference on Systems, Man and Cybernetics (SMC2007), pages 1670-1675, Montreal, Oct. 2007. IEEE, IEEE.
	In this paper, we study neural network (NN) based face detection. The main purpose is to reduce the complexity of the NN detector andthus speedup training and detection through linear dimensionalityreduction. Neither principle component analysis (PCA) nor lineardiscriminant analysis (LDA) is good for this purpose. PCA oftenreduces descriptive and discriminative information together. On theother hand, LDA maps all data into a (C-1)-Dimensional featurespace, where C=2 for face detection. Since face detection ishighly non-linear, classification in 1-dimensional space is clearlynot enough. In this paper, we propose a new method for facedetection. In this method, the problem is first changed to amulti-class problem using clustering. A modified LDA (m-LDA) isthen proposed to extract useful features. The NN is used to makethe final decision. Here, m-LDA is used to minimize thewithin-cluster variance between all face clusters; and maximizethe between-cluster variance between all face clusters and allnon-face data points. The feature space so obtained has adimensionality less than that of the original problem. To validatethe proposed method, we conducted several experiments with fourmethods, namely the proposed method, NN, PCA+NN, and LDA+NN. Resultsshow that the proposed method can provide lower false positive andfalse negative errors for test images.
[qf-zhao-03:2007]	C. F. Lv, Q. F. Zhao, and Z. W. Liu. A exible non-linear PCA encoder for still image compression. In Daming Wei, editor, Proc. 7th IEEE International Conference on Computer and Information Technology (CIT07), pages 645-650, Aizuwakamatsu, Japan, Oct. 2007. U-Aizu, IEEE.
	The main hindrance to develop a principal componentanalysis (PCA) encoder for image compression is the poorgeneralization ability of PCA. In this paper, we present a exiblesemi-universal image encoder based on the recently proposed non-linearPCA framework. Unlike other PCA techniques with a fixed order ofprincipal components, the proposed encoder can exibly determinewhich component is more significant to the quality of compressionaccording to the characteristics of the sub-image block to encode. Theproposed encoder is used to compress still gray level images, andexperimental results indicate that it can provide very goodgeneralization ability as well as high compression ratio.
[qf-zhao-04:2007]	A. Tamura and Q. F. Zhao. Rough Common Vector: A New Approach to Face Recognition. In IEEE, editor, Proc. IEEE International Conference on Systems, Man and Cybernetics (SMC2007), pages 2366-2371, Montreal, Oct. 2007. IEEE, IEEE.
	Face recognition is one of the most fundamental functions for manykinds of intelligent robots. Among many methods proposed for facerecognition, linear approaches such as Eigenface or principalcomponent analysis (PCA), Fisher's linear discriminant (FLD),Fisherface, and common vector (CV), have attracted great attentionbecause of their simplicity. It is known that when the number oftraining examples is large enough, we should use FLD; otherwise, weshould use CV. This paper proposes a rough common vector (RCV)approach. The basic idea of RCV is to divide the feature space intotwo subspaces. One is spanned by the eigenvectors corresponding tothe largest eigenvalues of the within-class scatter matrix, andanother is spanned by the eigenvectors corresponding to the smallesteigenvalues. The later plays the role of the null space of thewithin-class scatter matrix, and is important for extracting usefuldiscriminative features for recognition. RCV can be used regardlessthe within-class scatter matrix is singular or not. Experimentalresults on four databases show that RCV outperforms nearest neighborclassifier, Eigenface, Fisherface, and CV in most cases.
[qf-zhao-05:2007]	K. Sakamoto and Q. F. Zhao. Generating comprehensible moving policies for mobile robots through co-evolution of navigators and environment patterns. In IEEE, editor, Proc. IEEE International Conference on Systems, Man and Cybernetics (SMC07), pages 2455{2460, Montreal, Oct. 2007. IEEE, IEEE.
	To make robot navigators that has generalize well, It isimportant things that we should evaluate the navigators using as manyenvironment patterns as possible during evolution. If we use as fewenvironment patterns to reduce the computational cost, it is difficultto know in advance what environment patterns can evolve goodnavigators. To solve this problem, we study two differentapproaches. One is the co-evolutionary algorithm (CEA) that evolvesthe navigators and the environment patterns together, and the other isto select good environment patterns through incremental evolution(IE). Detailed considerations in using CEA and IE are described inthis paper, and their efficiency is compared with each other throughsimulations.
[yliu-01:2007]	Y. Liu. Evolutionary Programming with Operator Adaptation. In Pro- ceedings of the Seventh International Conference on Computer and Information Technology, pages 306-311. IEEE Computer Society, Oct. 2007.
	This paper investigated evolutionary programming with operator adaptation at both population level and individual level. The fitness distributions were employed to update operators at population level while the immediate reward or punishment from the feedback of mutations was applied to change operators at individual level. Experimental results had shown that long jump operators could actually have smaller average winning step sizes. Through observing the evolution of step sizes and fitness distribution values for each mutation operator, it was discovered that small-stepping operator could become the only dominant operator while other more capable operators with long jumps had only been applied at rather low probabilites.
[yliu-02:2007]	Y. Liu. Create Stable Neural Networks for Better Classification. In Proceedings of the 14th International Conference on Neural Information Processing (the proceedings will be published in a volume of Lecture Notes in Computer Science by Springer in 2008. Springer, Nov. 2007.
	This paper studies how to make a learned neural network more stable by letting it re-learn on those data points on which it disagrees with some other learned networks. Through such re-learning process, the re-learned neural network could be expected to have smaller difference or variance from one run to another. Therefore, for those problems where the performance of learned neural networks is greatly affected by their variance, this re-learning would be quite effective. Experimental results have been conducted on four real world problems to show the relationship between variance and performance in neural network learning, and explain how and when such re-learning works.
[yliu-03:2007]	Y. Liu. Operator Adaptation in Evolutionary Programming. In L. Kang, Y. Liu, and S. Zeng, editors, Intelligence Computation and Applications, Lecture Notes in Computer Science, Vol. 4683, pages 90-99. Springer, Sept. 2007.
	Operator adaptation in evolutionary programming is investigated from both population level and individual level in this paper. The updating rule for operator adaptation is defined based on the fitness distributions at population level compared to the immediate reward or punishment from the feedback of mutations at individual level. Through observing the behaviors of operator adaptation in evolutionary programming, it is discovered that a small-stepping operator could become a dominant operator when other operators have rather larger step sizes. Therefore, it is possible that operator adaptation could lead to slow evolution when operators are adapted freely by themselves.
[yliu-04:2007]	Y. Liu. Lead to Better Classication from Ensemble Learning. In Proceedings of the Seventh International Conference on Optimization: Techniques and Applications, pages 441-442. Universal Academy Press, Inc., Dec. 2007.
	This paper discuss how to make a learned neural network have a better classification by letting it re-learn from a trained ensemble. Such re-learning process would force the learned neural network to have similar input-output mapping of the ensemble so that the re-learned neural network would have smaller variance than it had before the re-learning. This re-learning process not only reduces the variance of learned neural networks, but also keeps the bias of those learned neural networks small. Therefore, the re-learning is able to decrease the sum of bias and variance of the learned neural networks, and lead the re-learned neural networks to have better performance. Experimental results on four real world problems were presented to show how the variance was decreased in the re-learning process while the bias was kept at small value.

Book

[yliu-05:2007]	L. Kang, Y. Liu, and S. Zeng. Evolvable Systems: From Biology to Hardware. Number 4684 in Lecture Notes in Computer Science. Springer, 2007.
[yliu-06:2007]	L. Kang, Y. Liu, and S. Zeng. Intelligence Computation and Applications. Number 4683 in Lecture Notes in Computer Science. Springer, 2007.

Academic Activities

[yliu-07:2007]	Y. Liu, 2006-2007. General Co-Chair of the 2nd International Symposium on Intelligence Computation and Applications (ISICA2007)br>
[yliu-08:2007]	Y. Liu, 2006-2007. Program Chair of the 7th International Conference on Evolvable Systems: From Biology to Hardware (ICES2007)

Ph.D, Master and Graduation Theses

[qf-zhao-06:2007]	Kazuhiko Hirakuri. Graduate Thesis: Head pose recognition with the NNC-Trees, University of Aizu, 2007. Thesis Advisor: Qiangfu Zhao
[qf-zhao-07:2007]	Shuuichi Takahashi. Graduation Thesis: Face detection based on combination of NN, clustering, and LDA, University of Aizu, 2007. Thesis Advisor: Qiangfu Zhao
[qf-zhao-08:2007]	Youhei Harada. Graduation Thesis: Interpreting fuzzy rules evolved for mobile robots, University of Aizu, 2007. Thesis Advisor: Qiangfu Zhao
[qf-zhao-09:2007]	Hiroyuki Saito. Graduation Thesis: A study on improving face recognition performance of NNC-Trees, University of Aizu, 2007. Thesis Advisor: Qiangfu Zhao
[qf-zhao-10:2007]	Hirotaka Takahashi. Graduation Thesis: A rough null space approach to face recognition, University of Aizu, 2007. Thesis Advisor: Qiangfu Zhao
[qf-zhao-11:2007]	Akira Saito. Graduation Thesis: Semi-automatic morphing based on image recognition, University of Aizu, 2007. Thesis Advisor: Qiangfu Zhao
[qf-zhao-12:2007]	Hiroyuki Nakamura. Master Thesis: Improvement of security strength for morphing based information hiding, University of Aizu, 2007. Thesis Advisor: Qiangfu Zhao
[qf-zhao-13:2007]	Ryo Watanabe. Master Thesis: Investigation of effectiveness of NN and NN ensemble in financial time-series prediction, University of Aizu, 2007. Thesis Advisor: Qiangfu Zhao
[qf-zhao-14:2007]	Kazuya Nozawa. Graduation Thesis: Spam mail detection with Bayesian filter, University of Aizu, 2007. Thesis Advisor: Qiangfu Zhao
[yliu-09:2007]	Taiga Watanabe. Evaluative Expression Classication Using Bayesian Filter, University of Aizu, 2007. Thesis Advisor: Liu, Y
[yliu-10:2007]	Masashi Takeda. Nikkei 225 Open Price Prediction Based on Neural Networks, University of Aizu, 2007. Thesis Advisor: Liu, Y

2009 THE UNIVERSITY OF AIZU ALL RIGHTS RESERVED