Let us consider a specific action or situation (called event) that takes place within a time series. The objective in early detection is to build a decision function that is able to go off as soon as possible from the onset of an occurrence of this event. This implies making a decision with an incomplete information. This paper proposes a novel framework that i) guarantees that a detection made with a partial observation will also occur at full observation of the time-series; ii) incorporates in a consistent manner the lack of knowledge about the minimal amount of information needed to make a decision. The proposed detector is based on mapping the temporal sequences to a landmarking space thanks to appropriately designed similarity functions. As a by-product, the framework benefits from a scalable training algorithm and a theoretical guarantee concerning its generalization ability. We also discuss an important improvement of our framework in which decision function can still be made reliable while being more expressive. Our experimental studies provide compelling results on toy data, presenting the trade-off that occurs when aiming at accuracy, earliness and reliability. Results on real physiological and video datasets show that our proposed approach is as accurate and early as state-of-the-art algorithm, while ensuring reliability and being far more efficient to learn.

This paper addresses the problem of automatic feature extraction for signal classification. In order to handle non-stationarity, features are designed in the time-frequency domain using a Filter Bank as the mapping function, which enables an easy interpretation for practitioners. The strategy adopted is to jointly learn a Filter Bank with a Support Vector Machine by casting the optimization program as a Multiple Kernel Learning problem. This solves the program for a finite set of filters. Thus, in order to handle an infinite number of filters, a novel active constraint algorithm is proposed based on the latest breakthroughs. Our method has been tested on a toy dataset and compared to classical methods with competitive results.

The rate-distortion performance of a distributed video coding system strongly depends on the characteristics of the side information. One could naïvely think that the best side information is the one with the largest PSNR with respect to the original corresponding image. However, previous works have shown that this is not always the case and a reduction of the side information MSE does not always translate into better rate-distortion performance for the complete system. The scope of this paper is to explore a set of metrics other than the PSNR and explicitly designed to classify the side information with respect to its impact on the end-to-end compression performance. A first contribution is to define an experimental framework that can be used to meaningfully compare different metrics for side information evaluation. As a second contribution, our analysis allows to understand why in some cases PSNR-based metrics provide a fairly reliable estimation of the side information quality, while in other cases they do not. This analysis also allows us to introduce a set of new metrics that are better adapted for side information effectiveness evaluation, and that are based on a suitable power of the absolute difference between side information and the original image, or on the Hamming distance between the respective transform coefficients. Besides their theoretical interest, these new metrics can also improve the rate-distortion performance of some distributed video coding systems such as the hash-based ones. We observe improvement up to 74% rate reduction in a simple study case.

The Laplacian model is the standard distribution for correlation noise estimation at the turbodecoder in Wyner-Ziv coding schemes. In practice, this hypothesis is not always satisfied and, regularly, the estimated model sensibly differs from the error distribution. In this work, we prove that using a model better fitted to the true distribution improves the performances, and we thus propose to use the more general exponential power distribution (EPD) which has never been tested in a distributed video coding context. Gains in rate-distortion over the Laplacian model are illustrated by results on several video sequences, showing that the EPD model outperforms the Laplacian one in off-line (oracle) as well as in on-line (practical implementation) modes. These results also indicate that, in some cases, the online EPD model reduces the bitrate even over the off-line Laplacian model.

When an oversampled FIR filter bank structure is used for signal analysis, a main problem is to guarantee its invertibility and to be able to determine an inverse synthesis filter bank. As the analysis scheme corresponds to a redundant decomposition, there is no unique inverse filter bank and some of the solutions can lead to artifacts in textured image filtering applications. In this paper, the flexibility in the choice of the inverse filter bank is exploited to find the best-localized impulse responses. The design is performed by solving a constrained optimization problem which is reformulated in a smaller dimensional space. Application to seismic data clearly shows the improvements brought by the optimization process.

In a previous work, we proposed a relatively simple method to build non separable perfect reconstruction oversampled lapped transforms. The main drawback of this method was that the redundancy factor was constrained to be equal to the overlapping one. This constitutes a strong limitation for applications such as seismic processing involving three-dimensional data sets. The memory requirements may indeed become hard to meet if the redundancy is not reduced. In this paper, we propose an approach to guarantee that a given lapped transform is invertible by a finite length filter bank. We show how to compute a corresponding synthesis filter bank. The proposed analysis/synthesis filter bank system is applied to directional filtering of noisy three-dimensional seismic data