Pub Date : 2002-11-07DOI: 10.1109/NNSP.2002.1030059
M. Aoki, K. Furuya
We propose a real-time source separation method that works well even under reverberant conditions. Previously, we proposed a method called SAFIA, which segregates sound sources by using sound localization cues acquired by multiple microphones. Under reverberant conditions, SAFIA suffers from "spectral overlap caused by reverberation", which introduces distortion into the separated speech signals. Extending the concept of SAFIA, we propose a new method (WAFD-SAFIA) based on simple signal-processing operations. WAFD-SAFIA significantly reduces the effects of "spectral overlap caused by reverberation". Computing the SNR (signal-to-noise ratio) and SDR (signal-to-distortion ratio) for both methods, we found that this new method outperformed SAFIA in a realistic environment. Moreover, to clarify the effect of frequency resolution on SAFIA, we determined whether a given frequency resolution decreased the overlap between the frequency components of two speech signals.
{"title":"Real-time source separation based on sound localization in a reverberant environment","authors":"M. Aoki, K. Furuya","doi":"10.1109/NNSP.2002.1030059","DOIUrl":"https://doi.org/10.1109/NNSP.2002.1030059","url":null,"abstract":"We propose a real-time source separation method that works well even under reverberant conditions. Previously, we proposed a method called SAFIA, which segregates sound sources by using sound localization cues acquired by multiple microphones. Under reverberant conditions, SAFIA suffers from \"spectral overlap caused by reverberation\", which introduces distortion into the separated speech signals. Extending the concept of SAFIA, we propose a new method (WAFD-SAFIA) based on simple signal-processing operations. WAFD-SAFIA significantly reduces the effects of \"spectral overlap caused by reverberation\". Computing the SNR (signal-to-noise ratio) and SDR (signal-to-distortion ratio) for both methods, we found that this new method outperformed SAFIA in a realistic environment. Moreover, to clarify the effect of frequency resolution on SAFIA, we determined whether a given frequency resolution decreased the overlap between the frequency components of two speech signals.","PeriodicalId":117945,"journal":{"name":"Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing","volume":"23 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125534822","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-07DOI: 10.1109/NNSP.2002.1030012
J. Kittler, K. Messer
We investigate two trainable methods of classifier fusion in the context of multimodal personal identity verification involving eight experts which exploit voice characteristics and frontal face biometrics. As baseline classifier combination methods, simple fusion rules (Sum and Vote) which do not require any training are used. The results of experiments on the XM2VTS database show that all four combination methods investigated yield improved performance. Trainable fusion strategies do not appear to offer better performance than simple rules.
{"title":"Fusion of multiple experts in multimodal biometric personal identity verification systems","authors":"J. Kittler, K. Messer","doi":"10.1109/NNSP.2002.1030012","DOIUrl":"https://doi.org/10.1109/NNSP.2002.1030012","url":null,"abstract":"We investigate two trainable methods of classifier fusion in the context of multimodal personal identity verification involving eight experts which exploit voice characteristics and frontal face biometrics. As baseline classifier combination methods, simple fusion rules (Sum and Vote) which do not require any training are used. The results of experiments on the XM2VTS database show that all four combination methods investigated yield improved performance. Trainable fusion strategies do not appear to offer better performance than simple rules.","PeriodicalId":117945,"journal":{"name":"Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing","volume":"349 2","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"113966976","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-07DOI: 10.1109/NNSP.2002.1030069
Zhiming Zhang, Seung-Gi Chang, Jeonghoon Park, Yongje Kim
A new SOLPN (self-organizing learning Petri net)-based rate control algorithm for an MPEG encoder is proposed. The idea is to use SOLPN to realize the RD (rate distortion) model, which is self-organized on line and adaptively updated frame by frame. The method does not require off-line pre-training; hence it is geared toward real-time coding. The comparative results on the examples suggest that our proposed rate control schemes encode video sequences with fewer frame skips, providing good subjective quality and higher PSNR, compared to VM18.
{"title":"A new SOLPN-based rate control algorithm for MPEG video coding","authors":"Zhiming Zhang, Seung-Gi Chang, Jeonghoon Park, Yongje Kim","doi":"10.1109/NNSP.2002.1030069","DOIUrl":"https://doi.org/10.1109/NNSP.2002.1030069","url":null,"abstract":"A new SOLPN (self-organizing learning Petri net)-based rate control algorithm for an MPEG encoder is proposed. The idea is to use SOLPN to realize the RD (rate distortion) model, which is self-organized on line and adaptively updated frame by frame. The method does not require off-line pre-training; hence it is geared toward real-time coding. The comparative results on the examples suggest that our proposed rate control schemes encode video sequences with fewer frame skips, providing good subjective quality and higher PSNR, compared to VM18.","PeriodicalId":117945,"journal":{"name":"Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing","volume":"136 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114314503","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-07DOI: 10.1109/NNSP.2002.1030022
G. Barreto, A. Araujo
We propose an unsupervised neural modelling technique, called vector-quantized temporal associative memory (VQTAM), which enables Kohonen's self-organizing map (SOM) to approximate nonlinear dynamical mappings globally. A theoretical analysis of the VQTAM scheme demonstrates that the approximation error decreases as the SOM training proceeds. The SOM is compared with standard MLP and RBF networks in the forward and inverse identification of a hydraulic actuator. The simulation results produced by the SOM are as accurate as those produced by the MLP network, and better than those produced by the RBF network; both the MLP and the RBF being supervised algorithms. The SOM is also less sensitive to weight initialization than MLP networks. The paper is concluded with a brief discussion on the main properties of the VQTAM approach.
{"title":"Temporal associative memory and function approximation with the self-organizing map","authors":"G. Barreto, A. Araujo","doi":"10.1109/NNSP.2002.1030022","DOIUrl":"https://doi.org/10.1109/NNSP.2002.1030022","url":null,"abstract":"We propose an unsupervised neural modelling technique, called vector-quantized temporal associative memory (VQTAM), which enables Kohonen's self-organizing map (SOM) to approximate nonlinear dynamical mappings globally. A theoretical analysis of the VQTAM scheme demonstrates that the approximation error decreases as the SOM training proceeds. The SOM is compared with standard MLP and RBF networks in the forward and inverse identification of a hydraulic actuator. The simulation results produced by the SOM are as accurate as those produced by the MLP network, and better than those produced by the RBF network; both the MLP and the RBF being supervised algorithms. The SOM is also less sensitive to weight initialization than MLP networks. The paper is concluded with a brief discussion on the main properties of the VQTAM approach.","PeriodicalId":117945,"journal":{"name":"Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing","volume":"119 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133551065","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-07DOI: 10.1109/NNSP.2002.1030026
A. Seghouane, Y. Moudden, G. Fleury
Injecting noise to the inputs during the training of feedforward neural networks (FNN) can improve their generalization performance remarkably. Reported works justify this fact arguing that noise injection is equivalent to a smoothing regularization with the input noise variance playing the role of the regularization parameter. The success of this approach depends on the appropriate choice of the input noise variance. However, it is often not known a priori if the degree of smoothness imposed on the FNN mapping is consistent with the unknown function to be approximated. In order to have a better control over this smoothing effect, a cost function putting in balance the smoothed fitting induced by the noise injection and the precision of approximation, is proposed. The second term, which aims at penalizing the undesirable effect of input noise injection or controlling the deviation of the random perturbed cost, was obtained by expressing a certain distance between the original cost function and its random perturbed version. In fact, this term can be derived in general for parametrical. models that satisfy the Lipschitz property. An example is included to illustrate the effectiveness of learning with this proposed cost function when noise injection is used.
{"title":"On learning feedforward neural networks with noise injection into inputs","authors":"A. Seghouane, Y. Moudden, G. Fleury","doi":"10.1109/NNSP.2002.1030026","DOIUrl":"https://doi.org/10.1109/NNSP.2002.1030026","url":null,"abstract":"Injecting noise to the inputs during the training of feedforward neural networks (FNN) can improve their generalization performance remarkably. Reported works justify this fact arguing that noise injection is equivalent to a smoothing regularization with the input noise variance playing the role of the regularization parameter. The success of this approach depends on the appropriate choice of the input noise variance. However, it is often not known a priori if the degree of smoothness imposed on the FNN mapping is consistent with the unknown function to be approximated. In order to have a better control over this smoothing effect, a cost function putting in balance the smoothed fitting induced by the noise injection and the precision of approximation, is proposed. The second term, which aims at penalizing the undesirable effect of input noise injection or controlling the deviation of the random perturbed cost, was obtained by expressing a certain distance between the original cost function and its random perturbed version. In fact, this term can be derived in general for parametrical. models that satisfy the Lipschitz property. An example is included to illustrate the effectiveness of learning with this proposed cost function when noise injection is used.","PeriodicalId":117945,"journal":{"name":"Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing","volume":"174 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126024927","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-07DOI: 10.1109/NNSP.2002.1030078
S. Vitabile, A. Gentile, G. B. Dammone, F. Sorbello
An automatic road sign recognition system, A(RS)/sup 2/, is aimed at the detection and recognition of one or more road signs from real-world color images. The authors have proposed an A(RS)/sup 2/ able to detect and extract sign regions from real world scenes on the basis of their color and shape features. Classification is then performed on extracted candidate regions using multi-layer perceptron neural networks. Although system performances are good in terms of both sign detection and classification rates, the entire process requires a large computational time, so real-time applications are not allowed. We present the implementation of the neural layer on the Georgia Institute of Technology SIMD (single instruction, multiple data) pixel processor. Experimental trials supporting the feasibility of real-time processing on this platform are also reported.
{"title":"Multi-layer perceptron mapping on a SIMD architecture","authors":"S. Vitabile, A. Gentile, G. B. Dammone, F. Sorbello","doi":"10.1109/NNSP.2002.1030078","DOIUrl":"https://doi.org/10.1109/NNSP.2002.1030078","url":null,"abstract":"An automatic road sign recognition system, A(RS)/sup 2/, is aimed at the detection and recognition of one or more road signs from real-world color images. The authors have proposed an A(RS)/sup 2/ able to detect and extract sign regions from real world scenes on the basis of their color and shape features. Classification is then performed on extracted candidate regions using multi-layer perceptron neural networks. Although system performances are good in terms of both sign detection and classification rates, the entire process requires a large computational time, so real-time applications are not allowed. We present the implementation of the neural layer on the Georgia Institute of Technology SIMD (single instruction, multiple data) pixel processor. Experimental trials supporting the feasibility of real-time processing on this platform are also reported.","PeriodicalId":117945,"journal":{"name":"Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing","volume":"4 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126148672","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-07DOI: 10.1109/NNSP.2002.1030049
N. Poh, Samy Bengio, J. Korczak
In this study, two techniques that can improve the authentication process are examined: (i) multiple samples and (ii) multiple biometric sources. We propose the fusion of multiple samples obtained from multiple biometric sources at the score level. By using the average operator, both the theoretical and empirical results show that integrating as many samples and as many biometric sources as possible can improve the overall reliability of the system. This strategy is called the multi-sample multi-source approach. This strategy was tested on a real-life database using neural networks trained in one-versus-all configuration.
{"title":"A multi-sample multi-source model for biometric authentication","authors":"N. Poh, Samy Bengio, J. Korczak","doi":"10.1109/NNSP.2002.1030049","DOIUrl":"https://doi.org/10.1109/NNSP.2002.1030049","url":null,"abstract":"In this study, two techniques that can improve the authentication process are examined: (i) multiple samples and (ii) multiple biometric sources. We propose the fusion of multiple samples obtained from multiple biometric sources at the score level. By using the average operator, both the theoretical and empirical results show that integrating as many samples and as many biometric sources as possible can improve the overall reliability of the system. This strategy is called the multi-sample multi-source approach. This strategy was tested on a real-life database using neural networks trained in one-versus-all configuration.","PeriodicalId":117945,"journal":{"name":"Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing","volume":"38 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127712429","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-07DOI: 10.1109/NNSP.2002.1030081
K. Hara, T. Omori, Reiko Ueno
This paper describes a model, based on a Markov process model, of daily human behavior in an intelligent house where human behavior is observed with small motion detectors. The number of sensor states is reduced to a few dozen by a vector quantization method, and transitions within this reduced set of states are observed. Then, the state transition probability and the transition duration time distribution are used as the templates of daily human activity. The validity of those templates is evaluated by detecting unusual human behavior in three sets of different human behavior data. Successful detection of unusual behavior without any a priori knowledge shows the effectiveness of probabilistic human behavior description in the intelligent house.
{"title":"Detection of unusual human behavior in intelligent house","authors":"K. Hara, T. Omori, Reiko Ueno","doi":"10.1109/NNSP.2002.1030081","DOIUrl":"https://doi.org/10.1109/NNSP.2002.1030081","url":null,"abstract":"This paper describes a model, based on a Markov process model, of daily human behavior in an intelligent house where human behavior is observed with small motion detectors. The number of sensor states is reduced to a few dozen by a vector quantization method, and transitions within this reduced set of states are observed. Then, the state transition probability and the transition duration time distribution are used as the templates of daily human activity. The validity of those templates is evaluated by detecting unusual human behavior in three sets of different human behavior data. Successful detection of unusual behavior without any a priori knowledge shows the effectiveness of probabilistic human behavior description in the intelligent house.","PeriodicalId":117945,"journal":{"name":"Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129677827","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-07DOI: 10.1109/NNSP.2002.1030018
R. Murphy, M. Velliste, G. Porreca
The ongoing biotechnology revolution promises a complete understanding of the mechanisms by which cells and tissues carry out their functions. Central to that goal is the determination of the function of each protein that is present in a given cell type, and determining a protein's location within cells is critical to understanding its function. As large amounts of data become available from genome-wide determination of protein subcellular location, automated approaches to categorizing and comparing location patterns are urgently needed. Since subcellular location is most often determined using fluorescence microscopy, we have developed automated systems for interpreting the resulting images. We report here improved numeric features for describing such images that are fairly robust to image intensity binning and spatial resolution. We validate these features by using them to train neural networks that accurately recognize all major subcellular patterns with an accuracy higher than previously reported. Having validated the features by using them for classification, we also demonstrate using them to create Subcellular Location Trees that group similar proteins and provide a systematic framework for describing subcellular location.
{"title":"Robust classification of subcellular location patterns in fluorescence microscope images","authors":"R. Murphy, M. Velliste, G. Porreca","doi":"10.1109/NNSP.2002.1030018","DOIUrl":"https://doi.org/10.1109/NNSP.2002.1030018","url":null,"abstract":"The ongoing biotechnology revolution promises a complete understanding of the mechanisms by which cells and tissues carry out their functions. Central to that goal is the determination of the function of each protein that is present in a given cell type, and determining a protein's location within cells is critical to understanding its function. As large amounts of data become available from genome-wide determination of protein subcellular location, automated approaches to categorizing and comparing location patterns are urgently needed. Since subcellular location is most often determined using fluorescence microscopy, we have developed automated systems for interpreting the resulting images. We report here improved numeric features for describing such images that are fairly robust to image intensity binning and spatial resolution. We validate these features by using them to train neural networks that accurately recognize all major subcellular patterns with an accuracy higher than previously reported. Having validated the features by using them for classification, we also demonstrate using them to create Subcellular Location Trees that group similar proteins and provide a systematic framework for describing subcellular location.","PeriodicalId":117945,"journal":{"name":"Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing","volume":"20 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123388631","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-07DOI: 10.1109/NNSP.2002.1030046
S. Sato, T. Umezaki
In this paper, we propose a segmentation method for identifying a fingerprint image with the variation of vertical length using a recurrent neural network (RNN). Group delay spectra and histograms of horizontal pixel line are used as input features fed into the RNN and two target output patterns with and without consideration of state dependency are introduced for learning. The method composed of the histogram learning and the state-dependent target indicates the best performance. When the tolerable segmentation error is 60 pixels, a segmentation rate of 97.2% is obtained. In comparison with the rule-based method, this method has an advantage of about 10%. Furthermore, we show that this method has a characteristic different from the rule-based method in regard to segmentation faults, and the learning with the state-dependent target is more effective than that without the dependency.
{"title":"A fingerprint segmentation method using a recurrent neural network","authors":"S. Sato, T. Umezaki","doi":"10.1109/NNSP.2002.1030046","DOIUrl":"https://doi.org/10.1109/NNSP.2002.1030046","url":null,"abstract":"In this paper, we propose a segmentation method for identifying a fingerprint image with the variation of vertical length using a recurrent neural network (RNN). Group delay spectra and histograms of horizontal pixel line are used as input features fed into the RNN and two target output patterns with and without consideration of state dependency are introduced for learning. The method composed of the histogram learning and the state-dependent target indicates the best performance. When the tolerable segmentation error is 60 pixels, a segmentation rate of 97.2% is obtained. In comparison with the rule-based method, this method has an advantage of about 10%. Furthermore, we show that this method has a characteristic different from the rule-based method in regard to segmentation faults, and the learning with the state-dependent target is more effective than that without the dependency.","PeriodicalId":117945,"journal":{"name":"Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing","volume":"3 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115382668","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：