SIGNAL AND DATA PROCESSING
Year:2010 | Volume:- | Issue:1 (SERIAL 13)|Papers Count:7

Feature transformation methods can be divided into two linear and nonlinear approaches. The main idea of kernel method is that if the input feature space is transformed nonlinearly to a high-dimensional space, then transformed space will become linearly separable. This separation can be obtained according to different criteria. In kernel linear discriminative analysis (KLDA), the criterion is more discrimination between features in the new space. On the contrary, the kernel principal component analysis (KPCA) is based on more feature orthogonalization in the mapped space. In this paper, as criterion, we propose to minimize the classification error in the space created by the kernel. We presented and formulated our method as the name of KMCE (kernel minimum classification error). Our experiments are performed on UCI data sets using the different classification methods and compared with conventional linear and kernel based feature transformation techniques. Results show that our method has a higher recognition rate than other mentioned methods in the case of distance based classifiers. In addition, the performance of KMCE is as well as other methods for statistical and decision tree based classification approaches. Also, we conducted some speech recognition experiments on Aurora2. Results indicate that KMCE outperforms other nonlinear feature transformation methods.

Most of the recent studies have tried to create diversity in primary results and then applied a consensus function over all the obtained results to combine the weak partitions. In this paper a clustering ensemble method is proposed which is based on a subset of primary clusters. The main idea behind this method is using more stable clusters in the ensemble. The stability is applied as a goodness measure of the clusters. The clusters which satisfy a threshold of this measure are selected to participate in the ensemble. For combining the chosen clusters, a co-association based consensus function is applied. A new EAC based method which is called Extended Evidence Accumulation Clustering, EEAC, is proposed for constructing the Co-association Matrix from the subset of clusters. The proposed method is evaluated on five different UCI repository data sets. The empirical studies show the significant improvement of the proposed method in comparison with other ones.

Nowadays, automatic analysis of music signals has gained a considerable importance due to the growing amount of music data found on the Web. Music genre classification is one of the interesting research areas in music information retrieval systems. In this paper several techniques were implemented and evaluated for music genre classification including feature extraction, feature selection and music genre modeling on a database of 8 different music genres containing Celtic, Classic, Classic Piano, Jazz, Metal, Persian Classic, Relaxing and Dance music. This database was gathered from several albums composed by different musicians. Short, middle and long term features were studied and finally only short and middle term features were used in our experiments. The long term features were discarded due to their low performance in music genre classification. Two modeling types of the music genres were evaluated. In the first type, only distribution of the feature vectors was used and in the second type, the ordering of the feature vectors was taken into account. Some modeling techniques such as ANN, GMM, Decision Tree and SVM were used individually and in a hierarchical approach. We proposed a taxonomy which classifies the music genres in a hierarchy where there are a small number of classes in the root and large number of classes in leaves. In fact, each class at the root of taxonomy contains one or more music genres and each genre is represented as a leaf at the bottom of the taxonomy. In addition, several classifiers were used simultaneously, in a way that each of them classifies the music genres individually.The decision is finally made using a voting algorithm. Besides, several short-term feature extraction techniques which have successfully been applied in speech recognition, music instrument classification and also music genre classification were studied and after analysis of the experimental results using statistical measures and different combinations of features, a near optimal feature vector was selected.

In this paper, a new algorithm is introduced for localization of multiple speakers in echoic environments. The origin of localization is based on combination of TDOA estimates of each source obtained by the BSS algorithm in the time domain. A new BSS algorithm is proposed which improves the quality and channel identification compared to a reference technique and also reduces the computational cost in some cases. To solve the global permutation ambiguity of BSS algorithms, speech features are used. Simulation results show the effectiveness of these features for solving the later problem.

Machine translation of English sentences faces a big problem when it deals with phrasal verbs. Phrasal verb is a common structure occurring in English as a combination of a verb and a preposition, a verb and an adverb, or a verb with both an adverb and a preposition. Meaning of a phrasal verb is not compositional. The second part of the phrasal verbs which often is a preposition is called particle. The process of detecting a preposition as a particle or as an attachment in a preposition phrase can be a challenging problem. In this paper, we present a method which uses a combination of linguistic heuristic rules with a probabilistic English parser to disambiguate the role of prepositions. The aim of this disambiguation is to correctly detect the phrasal verbs in English to Persian machine translation system. Experiments on a corpus containing 520 sentences show that the quality of phrasal verb recognition in this system grows up to 87%.

Name entity recognition (NER) is a system that can identify one or more kinds of names in a text and classify them into specified categories. These categories can be name of people, organizations, companies, places (country, city, street, etc.), time related to names (date and time), financial values, percentages, etc. Although during the past decade a lot of researches has been done on NER in different languages, but lack of a system with admissible performance in Farsi texts is quietly sensible. In this paper, the Corpus of Research Center of Intelligent Signal Processing has been used to create a Farsi NER. In our proposed NER system, there exist three stages: preprocessing, feature extraction and classification. To prepare a data set in the preprocessing stage, by using the part of speech (POS) feature, names are extracted from text and then infinitives, time related names, counting names, and numbers are removed from data. This gives a more balanced data set for learning and classification. In the feature extraction stage, N-gram is computed as feature, and four classifiers (linear, KNN, Bayesian, Neural Network) is learned in the classification stage. Because of lack of variety in the time related names and a few number of mixture of time related names with names in the other categories, an auxiliary list is used to identifying them. The results of research show, neural network have better performance (99%) in distinct between the names of places and people. In general, KNN and linear classifiers obtain 91% success based on F-measure scale in classifying the names of places and people and general names. In classifying the time related names, using an auxiliary list, based on an F-measure scale, a 96% success was obtained.

Methods for face recognition which are based on face structure are among techniques without supervision and produce unfavorable results in the presence of linear changes in images. PCA is a linear transform and a powerful tool for data analysis but does not produce good results for face recognition when there are non-linear changes resulting from changes in position, intensity and gesture in the face image. To overcome this problem, methods based on face features are used. Gabor filtering which can be considered as a feature based method can be used in these cases. This paper presents a new face recognition algorithm by combining PCA and Gabor filtering methods. After Gabor filtering of each face image, a number of images is produced. Then, mean of these images is calculated and PCA is applied to it. The resulted principal components are then used for face recognition. The presented algorithm has been applied to face images from YaleB and ORL databases under different conditions. Results show that the new algorithm performs better than PCA or Gabor filtering methods when they are applied to face images independently.