Bilge Gunsel

Incremental subspace learning via non-negative matrix factorization

In this paper we introduce an incremental non-negative matrix factorization (INMF) scheme in order to overcome the difficulties that conventional NMF has in online processing of large data sets. The proposed scheme enables incrementally updating its factors by reflecting the influence of each observation on the factorization appropriately. This is achieved via a weighted cost function which also allows controlling the memorylessness of the factorization. Unlike conventional NMF, with its incremental nature and weighted cost function the INMF scheme successfully utilizes adaptability to dynamic data content changes with a lower computational complexity. Test results reported for two video applications, namely background modeling in video surveillance and clustering, demonstrate that INMF is capable of online representing data content while reducing dimension significantly.

Content-based access to art paintings

This paper introduces ArtHistorian, a content-based classification and indexing system that represents the visual content of art paintings by a six-dimensional feature set. The introduced feature set is robust to scale changes and can handle variations in lighting conditions. A nonlinear SVM classifier included in the system learns the characteristics of fundamental art movements and painting styles. A hybrid classifier that combines PCA representation of paintings with the SVM classification is also exploited. It is shown that ArtHistorian is capable of classifying art paintings based on painters as well as art movements with an accuracy of greater than 90% and its false alarm ratio is very small. The developed system enables the user to run content-based queries and to retrieve from painting databases created in XML format.

Reconstruction and Boundary Detection of Range and Intensity Images Using Multiscale MRF Representations

The basic difficulty encountered in filtering-based multiscale boundary detection methods is the elimination of noise and insignificant edges without distorting the shape of boundaries. These methods remove noise and unnecessary detail by blurring the input image at different scales, which results in the loss of positional accuracy at the image discontinuities. In this paper, a nonlinear multiscale boundary detection method which prevents the conflict between the detection and localization goals is introduced. The method uses multiscale representations of coupled Markov random fields and applies a stochastic regularization scheme based on the Bayesian approach. This allows the robust integration of boundary information extracted at multiple scales simultaneously. The scheme is applicable to intensity and range images as well as to sparse data and eliminates the dependency on edge operator size which is the main difficulty in filtering-based multiscale techniques.

Video Content Representation by Incremental Non-Negative Matrix Factorization

Nonnegative matrix factorization (NMF) is a powerful decomposition tool which has been used in several content representation applications recently. However, there are some difficulties in implementing NMF in on-line video applications. This paper introduces an incremental NMF (INMF) without deviating from conventional NMFs main objective function, which is minimizing the reconstruction error. The proposed algorithm is capable of modeling dynamic content of the video; thus controls contribution of the subsequent observations to the NMF representation properly. It is shown that the INMF preserves additive, parts-based representation capability of the NMF with a low computational load while offering dimension reduction. Experimental results are given to compare the reconstruction performances of the conventional and incremental NMF. In addition, video scene change detection and dynamic video content representation by INMF are investigated. Test results demonstrate that the INMF can be used as a powerful on-line factorization tool.

Robust Audio Watermark Decoding by Supervised Learning

Most of the watermark (WM) decoding schemes use correlation-based methods because of their simplicity. In these methods, the WM signal embedded through a secret key is assumed as uncorrelated with the host signal. This is a hard restriction that can never be achieved and correlation between the received signal and the secret key becomes greater than zero even though the received signal is un-watermarked. Mostly a decision threshold specified semi-automatically is used at the decoding site. Since the audio watermarking is a nonlinear process that guarantees the inaudibility, there is no analytic way of determining an optimal threshold value that makes the WM decoding problem harder. This paper introduces a learning scheme followed by a nonlinear classification thus eliminates the threshold specification problem. The decoding process is modelled as a three-class classification problem and support vector machines (SVMs) are used in the learning of the embedded data. The decoding and detection performances of the developed system are greater than 98% and 95%, respectively. When the watermark-to-signal-ratio (WSR) is higher than -30 dB, system false alarm ratios remain less than 2%. It is shown that the introduced WM decoding method is robust to additive noise and most of add/remove and filter attacks of Stirmark

Multiple model target tracking with variable rate particle filters

Article history: Available online 10 January 2012

A novel perceptual feature set for audio emotion recognition

We present a novel system for audio emotion recognition based on the Perceptual Evaluation of Audio Quality (PEAQ) model as described by the standard, ITU-R BS.1387–1 which provides a mathematical model resembling the human auditory system. The introduced feature set performs perceptual analysis in time, spectral and Bark domains thus enabling us to represent the statistics of emotional audio for arousal and valence modes with a small number of features. Unlike the existing systems, the proposed feature set learns statistical characteristic of emotional differences hence does not require data normalization to eliminate speaker or corpus dependency. Recognition performance obtained for the well known VAM and EMO-DB corpora show that the classification accuracy achieved by the proposed feature set outperforms the reported benchmarking results particularly for valence both for natural and acted emotional data.

Annealed SMC Samplers for Nonparametric Bayesian Mixture Models

We develop a novel online algorithm for posterior inference in Dirichlet Process Mixtures (DPM). Our method is based on the Sequential Monte Carlo (SMC) samplers framework that generalizes sequential importance sampling approaches. Unlike the existing methods, the framework enables us to retrospectively update long trajectories in the light of recent observations and this leads to sophisticated clustering update schemes and annealing strategies that seem to prevent the algorithm to get stuck around a local mode. The performance has been evaluated on a Bayesian Gaussian density estimation problem with an unknown number of mixture components. Our simulations suggest that the proposed annealing strategy outperforms conventional samplers. It also provides significantly smaller Monte Carlo standard error with respect to particle filtering given comparable computational resources.

Perceptually enhanced blind single-channel music source separation by Non-negative Matrix Factorization

This paper proposes a 2D Non-negative Matrix Factorization (NMF) based single-channel source separation algorithm that emphasizes perceptually important components of audio. Unlike the existing methods, the proposed scheme performs a psychoacoustic pre-processing on the mixture spectrogram in order to supress audio components that are not critical to human hearing sensation while amplifying the perceptually important ones. This yields the auditory spectrogram referred as sonogram of the observed audio mixture and the individual sources are then extracted by 2D NMF. Test results reported in terms of Signal-to-Distortion-Ratio (SDR), Signal-to-Inference-Ratio (SIR) and Signal-to-Artifact-Ratio (SAR) show that the proposed perceptually enhanced separation improves the quality of decomposed audio sources by 1.5-6.5 dB with a reduced computational complexity.

An integrated decoding framework for audio watermark extraction

This paper proposes a blind audio watermark extraction technique that allows performing watermark decoding while installing data synchronization. The proposed decoding algorithm employs correlation techniques supported by a wavelet denoising process, thus improves the decoding performance significantly. A data adaptive nonlinear MPEG Layer 1 Model 1 compatible watermark encoder is designed for watermark embedding. A channel encoder is also included into the system to take the advantage of error correction. The method does not require the original audio for decoding and it is robust to channel noise, filtering as well as stereo-to-mono conversions. It allows working at very low watermark-to-signal ratios thus preserves inaudibility.