Semi-supervised Extraction of Audio-Visual Sources

This report presents a semi-supervised method to jointly extract audio-visual sources from a scene. It consist of applying a supervised method to segment the video signal followed by an automatic process to properly separate the audio track. This approach starts with the user interaction to select the audio-visual target we want to cut. This labels the problem into two parts. One is the foreground, the object we aim to extract. The background is the remaining content which sets the other part. The segmentation method will be performed over the video signal to split its visual content into foreground and background. It is based on describing the video information as a volume. It is composed of pixels and relationships among them which set the features of the problem. This 3D structure can be interpreted as a graph. Therefore the segmentation is achieved by applying a Graph Cuts algorithm which provides the suitable results. In this work the audio separation process is an automatic task. Due to this we need to extract prior information from the audio-visual data available. Therefore a link between audio and video channels is deﬁned by an audio-visual motion map. This is a video sequence which provides the amount of synchronous motion with the audio track. An audio-based video diﬀusion method have been developed to obtain this kind of signal. Applying the video segmentation labelling to the audio-visual motion map we obtain the distribution of the amount motion into foreground and background. According to the range of these values it is possible to asses when the corresponding audio is on/oﬀ. At this point, audio samples of sources can be extract when the audio activity is due to only one. This is the prior information to learn audio models by means of spectral GMM estimation to assess the audio signals. Finally the separation method is applied. It consist of looking for the more suitable couple of states given the mixture spectrum of the total audio track. Therefore we achieve the audio separation goal. In this work we will study in detail all the stages towards the ﬁnal aim of our method, analyzing the performances of our approach.

Audio-Visual Fusion

Anna Llagostera Casanovas

The perception that we have about the world is influenced by elements of diverse nature. Indeed humans tend to integrate information coming from different sensory modalities to better understand their environment. Following this observation, scientists have been trying to combine different research domains. In particular, in joint audio-visual signal processing the information recorded with one or more video-cameras and one or more microphones is combined in order to extract more knowledge about a given scene than when analyzing each modality separately. In this thesis we attempt the fusion of audio and video modalities when considering one video-camera and one microphone. This is the most common configuration in electronic devices such as laptops and cellphones, and it does not require controlled environments such as previously prepared meeting rooms. Even though numerous approaches have been proposed in the last decade, the fusion of audio and video modalities is still an open problem. All the methods in this domain are based on an assumption of synchrony between related events in audio and video channels, i.e. the appearance of a sound is approximately synchronous with the movement of the image structure that has generated it. However, most approaches do not exploit the spatio-temporal consistency that characterizes video signals and, as a result, they assess the synchrony between single pixels and the soundtrack. The results that they obtain are thus sensitive to noise and the coherence between neighboring pixels is not ensured. This thesis presents two novel audio-visual fusion methods which follow completely different strategies to evaluate the synchrony between moving image structures and sounds. Each fusion method is successfully demonstrated on a different application in this domain. Our first audio-visual fusion approach is focused on the modeling of audio and video signals. We propose to decompose each modality into a small set of functions representing the structures that are inherent in the signals. The audio signal is decomposed into a set of atoms representing concentrations of energy in the spectrogram (sounds) and the video signal is concisely represented by a set of image structures evolving through time, i.e. changing their location, size or orientation. As a result, meaningful features can be easily defined for each modality, as the presence of a sound and the movement of a salient image structure. Finally, the fusion step simply evaluates the co-occurrence of these relevant events. This approach is applied to the blind detection and separation of the audio-visual sources that are present in a scene. In contrast, the second method that we propose uses basic features and it is more focused on the fusion strategy that combines them. This approach is based on a nonlinear diffusion procedure that progressively erodes a video sequence and converts it into an audio-visual video sequence, where only the information that is required in applications in the joint audio-visual domain is kept. For this purpose we define a diffusion coefficient that depends on the synchrony between video motion and audio energy and preserves regions moving coherently with the presence of sounds. Thus, the regions that are least diffused are likely to be part of the video modality of the audio-visual source, and the application of this fusion method to the unsupervised extraction of audio-visual objects is straightforward. Unlike many methods in this domain which are specific to speakers, the fusion methods that we present in this thesis are completely general and they can be applied to all kind of audio-visual sources. Furthermore, our analysis is not limited to one source at a time, i.e. all applications can deal with multiple simultaneous sources. Finally, this thesis tackles the audio-visual fusion problem from a novel perspective, by proposing creative fusion methods and techniques borrowed from other domains such as the blind source separation, nonlinear diffusion based on partial differential equations (PDE) and graph cut segmentation.

EPFL2011

Multimodal feature extraction and fusion for audio-visual speech recognition

Mihai Gurban

Multimodal signal processing analyzes a physical phenomenon through several types of measures, or modalities. This leads to the extraction of higher-quality and more reliable information than that obtained from single-modality signals. The advantage is two-fold. First, as the modalities are usually complementary, the end-result of multimodal processing is more informative than for each of the modalities individually, which represents the first advantage. This is true in all application domains: human-machine interaction, multimodal identification or multimodal image processing. The second advantage is that, as modalities are not always reliable, it is possible, when one modality becomes corrupted, to extract the missing information from the other one. There are two essential challenges in multimodal signal processing. First, the features used from each modality need to be as relevant and as few as possible. The fact that multimodal systems have to process more than just one modality means that they can run into errors caused by the curse of dimensionality much more easily than mono-modal ones. The curse of dimensionality is a term used essentially to say that the number of equally-distributed samples required to cover a region of space grows exponentially with the dimensionality of the space. This has important implications in the classification domain, since accurate models can only be obtained if an adequate number of samples is available, and obviously this required number of samples grows with the dimensionality of the features. Dimensionality reduction is thus a necessary step in any application dealing with complex signals, and this is achieved through selection, transforms or the combination of the two. The second essential challenge is multimodal integration. Since the signals involved do not necessarily have the same data rate, range or even dimensionality, combining information coming from such different sources is not straightforward. This can be done at different levels, starting from the basic signal level by combining the signals themselves, if they are compatible, up to the highest decision level, where only the individual decisions taken based on the signals are combined. Ideally, the fusion method should allow temporal variations in the relative importance of the two streams, to account for possible changes in their quality. However, this can only be done with methods operating at a high decision level. The aim of this thesis is to offer solutions to both these challenges, in the context of audio-visual speech recognition and speaker localization. Both these applications are from the field of human-machine interaction. Audio-visual speech recognition aims to improve the accuracy of speech recognizers by augmenting the audio with information extracted from the video, more particularly, the movement of the speaker's lips. This works well especially when the audio is corrupted, leading in this case to significant gains in accuracy. Speaker localization means detecting who is the active speaker in a audio-video sequence containing several persons, something that is useful for videoconferencing and the automated annotation of meetings. These two applications are the context in which we present our solutions to both feature selection and multimodal integration. First, we show how informative features can be extracted from the visual modality, using an information-theoretic framework which gives us a quantitative measure of the relevance of individual features. We also prove that reducing redundancy between these features is important for avoiding the curse of dimensionality and improving recognition results. The methods that we present are novel in the field of audio-visual speech recognition and we found that their use leads to significant improvements compared to the state of the art. Second, we present a method of multimodal fusion at the level of intermediate decisions using a weight for each of the streams. The weights are adaptive, changing according to the estimated reliability of each stream. This makes the system tolerant to changes in the quality of either stream, and even to the temporary interruption of one of the streams. The reliability estimate is based on the entropy of the posterior probability distributions of each stream at the intermediate decision level. Our results are superior to those obtained with a state of the art method based on maximizing the same posteriors. Moreover, we analyze the effect of a constraint typically imposed on stream weights in the literature, the constraint that they should sum to one. Our results show that removing this constraint can lead to improvements in recognition accuracy. Finally, we develop a method for audio-visual speaker localization, based on the correlation between audio energy and the movement of the speaker's lips. Our method is based on a joint probability model of the audio and video which is used to build a likelihood map showing the likely positions of the speaker's mouth. We show that our novel method performs better than a similar method from the literature. In conclusion, we analyze two different challenges of multimodal signal processing for two audio-visual problems, and offer innovative approaches for solving them.

EPFL2009

Blind Audio-Visual Source Separation based on Sparse Redundant Representations

Rémi Gribonval, Anna Llagostera Casanovas, Gianluca Monaci, Pierre Vandergheynst

In this paper we propose a novel method which is able to detect and separate audio-visual sources present in a scene. Our method exploits the correlation between the video signal captured with a camera and a synchronously recorded one-microphone audio track. In a ﬁrst stage, audio and video modalities are decomposed into relevant basic structures using redundant representations. Next, synchrony between relevant events in audio and video modalities is quantiﬁed. Based on this co-occurrence measure, audio-visual sources are counted and located in the image using a robust clustering algorithm that groups video structures exhibiting strong correlations with the audio. Next periods where each source is active alone are determined and used to build spectral Gaussian Mixture Models (GMMs) characterizing the sources acoustic behavior. Finally, these models are used to separate the audio signal in periods during which several sources are mixed. The proposed approach has been extensively tested on synthetic and natural sequences composed of speakers and music instruments. Results show that the proposed method is able to successfully detect, localize, separate and reconstruct present audio-visual sources.