Nikoletta Bassiou

Color image histogram equalization by absolute discounting back-off

A novel color image histogram equalization approach is proposed that exploits the correlation between color components and it is enhanced by a multi-level smoothing technique borrowed from statistical language engineering. Multi-level smoothing aims at dealing efficiently with the problem of unseen color values, either considered independently or in combination with others. It is applied here to the HSI color space for the probability of intensity and the probability of saturation given the intensity, while the hue is left unchanged. Moreover, the proposed approach is extended by an empirical technique, which is based on a hue preserving non-linear transformation, in order to eliminate the gamut problem. This is the second method proposed in the paper. The equalized images by the two methods are compared to those produced by other well-known methods. The better quality of the images equalized by the proposed methods is judged in terms of their visual appeal and objective figures of merit, such as the entropy and the Kullback-Leibler divergence estimates between the resulting color histogram and the multivariate uniform probability density function.

Online PLSA: Batch Updating Techniques Including Out-of-Vocabulary Words

A novel method is proposed for updating an already trained asymmetric and symmetric probabilistic latent semantic analysis (PLSA) model within the context of a varying document stream. The proposed method is coined online PLSA (oPLSA). The oPLSA employs a fixed-size moving window over a document stream to incorporate new documents and at the same time to discard old ones (i.e., documents that fall outside the scope of the window). In addition, the oPLSA assimilates new words that had not been previously seen (out-of-vocabulary words), and discards the words that exclusively appear in the documents to be thrown away. To handle the new words, Good-Turing estimates for the probabilities of unseen words are exploited. The experimental results demonstrate the superiority in terms of accuracy of the oPLSA over well known PLSA updating methods, such as the PLSA folding-in (PLSA fold.), the PLSA rerun from the breakpoint, the quasi-Bayes PLSA, and the Incremental PLSA. A comparison with respect to the CPU run time reveals that the oPLSA is the second fastest method after the PLSA fold. However, the better accuracy of the oPLSA than that of the PLSA fold. pays off the longer computation time. The oPLSA and the other PLSA updating methods together with online LDA are tested for document clustering and F1 scores are also reported.

Long distance bigram models applied to word clustering

Two novel word clustering techniques are proposed which employ long distance bigram language models. The first technique is built on a hierarchical clustering algorithm and minimizes the sum of Mahalanobis distances of all words after a cluster merger from the centroid of the class created by merging. The second technique resorts to probabilistic latent semantic analysis (PLSA). Next, interpolated long distance bigrams are considered in the context of the aforementioned clustering techniques. Experiments conducted on the English Gigaword corpus (second edition) demonstrate that: (1) the long distance bigrams, when employed in the two clustering techniques under study, yield word clusters of better quality than the baseline bigrams; (2) the interpolated long distance bigrams outperform the long distance bigrams in the same respect; (3) the long distance bigrams perform better than the bigrams, which incorporate trigger-pairs selected at various distances; and (4) the best word clustering is achieved by the PLSA that employs interpolated long distance bigrams. Both proposed techniques outperform spectral clustering based on k-means. To assess objectively the quality of the created clusters, relative cluster validity indices are estimated as well as the average cluster sense precision, the average cluster sense recall, and the F-measure are computed by exploiting ground truth extracted from the WordNet.

Speaker Diarization Exploiting the Eigengap Criterion and Cluster Ensembles

A novel system for speaker diarization is proposed that combines the eigengap criterion and cluster ensembles. No explicit assumptions on the number of speakers are made. Two variants of the system are developed. The first variant does not cluster the speech segments that are detected as outliers, while the second one does. The aforementioned system variants are assessed with respect to various metrics, such as the overall classification error, the average cluster purity, and the average speaker purity. Experiments are conducted on two-person dialogue scenes in movies as well as on news broadcasts from MDE RT-03 Training Data Speech Corpus released by the U.S. National Institute of Standards and Technology. In the latter case, the diarization error rate is also reported. It is demonstrated that the clustering performance does not degrade when outliers are present. Moreover, thanks to the eigengap criterion, the evaluation metrics are improved.

Frontal face detection using support vector machines and back-propagation neural networks

Face detection is a key problem in building systems that perform face recognition/verification and model-based image coding. Two algorithms for face detection that employ either support vector machines or backpropagation feedforward neural networks are described, and their performance is tested on the same frontal face database using the false acceptance and false rejection rates as quantitative figures of merit. The aforementioned algorithms can replace the explicitly-defined knowledge for facial regions and facial features in mosaic-based face detection algorithms.

RPLSA: A novel updating scheme for Probabilistic Latent Semantic Analysis

A novel updating method for Probabilistic Latent Semantic Analysis (PLSA), called Recursive PLSA (RPLSA), is proposed. The updating of conditional probabilities is derived from first principles for both the asymmetric and the symmetric PLSA formulations. The performance of RPLSA for both formulations is compared to that of the PLSA folding-in, the PLSA rerun from the breakpoint, and well-known LSA updating methods, such as the singular value decomposition (SVD) folding-in and the SVD-updating. The experimental results demonstrate that the RPLSA outperforms the other updating methods under study with respect to the maximization of the average log-likelihood and the minimization of the average absolute error between the probabilities estimated by the updating methods and those derived by applying the non-adaptive PLSA from scratch. A comparison in terms of CPU run time is conducted as well. Finally, in document clustering using the Adjusted Rand index, it is demonstrated that the clusters generated by the RPLSA are: (a) similar to those generated by the PLSA applied from scratch; (b) closer to the ground truth than those created by the other PLSA or LSA updating methods.

Greek folk music classification into two genres using lyrics and audio via canonical correlation analysis

We are interested in Greek folk music genre classification by resorting to canonical correlation analysis (CCA). Here, the genre is related to the place of origin of the song. The CCA learns a linear transformation of the song lyrics descriptors that is highly correlated with their genre labels as well as another linear transformation of the audio features extracted from music recordings, which is maximally correlated with their genre labels. In the latter task, thanks to the deep CCA (DCCA), deep nonlinear transformations of the audio features are learnt, which are maximally correlated with the genre labels. Experimental findings are disclosed for a two-class genre recognition problem, employing folk songs originated from Pontus and Asia Minor. It is demonstrated that the CCA achieves an average accuracy of 97.02% across the 5 folds, when the term frequency-inverse document frequency features model the song lyrics. By modeling the music signal of each song with 28 mel-frequency cepstral coefficients (MFCCs) extracted from each frame and averaged over all frames, the average accuracy of the CCA drops to 72.9% across the 5 folds. The DCCA yields an accuracy of 69% for audio-based genre recognition.

Symmetric α-stable sparse linear regression for musical audio denoising

A new musical audio denoising technique is proposed, when the noise is modeled by an α-stable distribution. The proposed technique is based on sparse linear regression with structured priors and uses Markov Chain Monte Carlo inference to estimate the clean signal model parameters and the α-stable noise model parameters. Experiments on noisy Greek folk music excerpts demonstrate better denoising for the α-stable noise assumption than the Gaussian white noise one.

Interpolated distanced bigram language models for robust word clustering

Summary form only given. Two methods for interpolating the distanced bigram language model are examined which take into account pairs of words that appear at varying distances within a context. The language models under study yield a lower perplexity than the baseline bigram model. A word clustering algorithm based on mutual information with robust estimates of the mean vector and the covariance matrix is employed in the proposed interpolated language model. The word clusters obtained by using the aforementioned language model are proved more meaningful than the word clusters derived using the baseline bigram.

Greek folk music classification using auditory cortical representations

In this paper, we deal with the classification of Greek folk songs into 8 classes associated with the region of origin of the songs. Motivated by the way the sound is perceived by the human auditory system, auditory cortical representations are extracted from the music recordings. Moreover, deep canonical correlation analysis (DCCA) is applied to the auditory cortical representations for dimensionality reduction. To classify the music recordings, either support vector machines (SVMs) or classifiers based on canonical correlation are employed. An average classification rate of 73.25 % is measured on a dataset of Greek folk songs from 8 regions, when the auditory cortical representations are classified by the SVMs. It is also demonstrated that the reduced features extracted by the DCCA yield an encouraging average classification rate of 66.27%. The latter features are shown to possess good discriminating properties.