First wall protection is a critical challenge in fusion machine operation, requiring rapid and precise diagnostic capabilities. The infrared diagnostics of long-pulse fusion machines, generate a large amount of video data. At WEST, the analysis of infrared thermal data is currently performed by Plasma-facing component Protection Officers (PPOs). This work presents a novel multimodal Large Language Model (LLM) approach to support PPOs in their diagnostic tasks. The framework addresses multiple challenges in fusion machine first wall monitoring by integrating infrared data (images and temperature curves) with domain-specific expertise. Key features include the preservation of data privacy, capturing a priori expert knowledge, quick post-pulse analysis capabilities, and the incorporation of expert feedback loops. The system employs prompt engineering and Low-Rank fine-tuning to embed domain knowledge directly into the diagnostic process. An ablation study using synthetic data demonstrated that a false recognition rate below 0.01 for hot spot characteristics could be achieved with 100 training samples. Based on these findings, the model was fine-tuned using 150 pairs of real-world images and corresponding prompts. This tool, once appropriately fine-tuned, was successfully implemented in the WEST control room during the C10 (10/2024 - 12/2024). Performance evaluation showed that of all the feedback collected, 32% were rated as ”OK” and 41% as ”positive response”. This development marks a significant advancement in automated diagnostic capabilities for fusion operations.

Research in sport sciences needs realistic simulators to elicit natural behaviors and understand the information-movement coupling underlying performance. This article outlines the theoretical and methodological considerations for optimizing the use of artificial intelligence (AI) in the development of a virtual reality setup (VR) for sport science use. The use case focused on the study information-movement coupling in a dyadic basket-throwing situation. An optimized software pipeline combining Google's MediaPipe SSD networks to capture the movement of a real attacker attempting to score in a virtual basket, and LSTM networks to classify the attacker's intentions and “intelligently” select from a library the animation of the defender's avatar was developed. As a result, the defender's avatar behaved on the basis of the behavior of a real attacker to contest the throw only if the attacker performed a guenuine throw. Measurements of two contradictory criteria (recognition rate of the attacker's intention and computation time) allowed to validate the optimized pipeline. The perspectives in the improvement of the simulator's realism, but also for theoretical, methodological, and applied viewpoints are discussed.

There are two main contributions in this paper. Firstly, we estimate the rank for the truncation of the Parafac decomposition in an optimal sense. For this, we propose a least squares criterion and justify the choice of the fast Nelder-Mead method to minimize this criterion. Secondly, we combine the truncation of the Parafac decomposition with multidimensional wavelet packet transform. A single rank value is estimated for each decomposition level, which simplifies the implementation. We exemplify the proposed method with an application to multispectral image denoising: we study the performance of the proposed method based on Parafac decomposition, compared to ForWaRD.

In this paper, we consider an issue of hand posture classification. We improve a recently proposed signature, a matrix containing the distance of all contour pixels to an arbitrary reference point. Adequate pre-processings ensure the invariance properties of the signature. Candidate postures are pre-selected with a surface criterion, and Principal Component Analysis (PCA) reduces the dimensionality of the data, which improves the classification process.

A nondestructive control system is proposed. We consider an application of defect detection in metal workpieces. To solve this defect detection problem, we propose an image processing system including denoising, segmentation, and mapping. A contour preserving denoising method based on wavelet decomposition is adapted; various tools of contour based image segmentation are combined adequately, to segment successively high contrast and low contrast contours; generalized Hough transform is adapted to map processed and reference workpieces without defects. Comparative results are proposed which involve the proposed image processing system, and two region based segmentation methods.

Subspace-based methods rely on dominant element selection from second order statistics. They have been extended to tensor processing, in particular to tensor data filtering. For this, the processed tensor is flattened along each mode successively, and singular value decomposition of the flattened matrix is classically performed. Data projection on the dominant singular vectors results in noise reduction. The numerical cost of SVD is elevated. Now, tensor processing methods include an ALS (Alternating Least Squares) loop, which implies that a large number of SVDs are performed. Fixed point algorithm estimates an a priori fixed number of singular vectors from a matrix. In this paper, we generalize fixed point algorithm as a higher-order fixed point algorithm to the estimation of only the required dominant singular vectors in a tensor processing framework. We compare the proposed method in terms of denoising quality and speed through an application to color image and hyperspectral image denoising.

Subspace-based methods rely on the selection of leading eigenvectors, associated with dominant eigenvalues. They have been extended to tensor data processing, such as denoising. Usually EVD (eigenvalue decomposition) is performed and data projection on leading eigenvectors results in noise reduction. Tensor processing methods, in particular multiway Wiener filtering algorithm, include an ALS (alternating least squares) loop, which involves several EVDs. Fixed point algorithm is a faster method than EVD to estimate a fixed number of eigenvectors. In this paper, we adapt fixed point algorithm to the estimation of only the required leading eigenvectors in a tensor processing framework. We adapt inverse power method to estimate the required noise variance. We provide a comparative study in terms of speed through an application to hyperspectral image denoising.

Thanks to a specific formalism for signal generation, it is possible to transpose an image processing problem to an array processing problem. For straight line characterization, the existing method subspace-based line detection (SLIDE) works on virtual signals generated on a linear antenna. In this paper we propose to retrieve circular and nearly circular contours in images. We propose a novel method for radius estimation, and we extend the estimation of circles to the retrieval of circular-like distorted contours. For this purpose we develop a new model for virtual signal generation: we simulate a circular antenna, so that a high resolution method can be employed for radius estimation. An application to biomedical imaging is proposed.

An important cause of performance loss in source localization in underwater acoustics is that towed flexible antennas deviate from the assumed rectilinear shape. In this work, the localization of sources in presence of phase errors is studied. Cancellation of phase errors is necessary to solve the source-localization problem. A previous work led to interesting results for antennas composed of a few sensors. We propose here a novel algorithm which is adapted to the antennas composed of a large number of sensors, keeping a small computational load. Our method is based on an orthogonality property between signal and noise subspaces and a novel optimization method: the robust DIviding RECTangles algorithm accelerated by spline interpolation. The performances of the proposed method are illustrated by applying it to the characterization of three sources.

Circular features are commonly sought in digital image processing. The subspace-based line detection (SLIDE) method proposed to estimate the center and the radius of a single circle. In this paper, we introduce a novel method for estimating several radii while extending the circle estimation to retrieve circular-like distorted contours. Particularly, we develop and validate a new model for virtual signal generation by simulating a circular antenna. The circle center is estimated by the SLIDE method. A variable speed propagation scheme toward the circular antenna yields a linear phase signal. Therefore, a high-resolution method provides the radius. Either the gradient method or the more robust combination of dividing rectangles and spline interpolation can extend this method extend this method for free form object segmentation. The retrieval of multiple non concentric circles and rotated ellipses is also considered. To evaluate the performance of the proposed methods, we compare them with a least-squares method, Hough transform, and gradient vector flow. We apply the proposed method to hand-made images while considering some real-world images.

Distorted curves retrieval is faced for robotic way screening, particle trajectory characterization, aerial and satellite image analysis. This image processing problem has been transposed to an array processing problem by adopting specific conventions. Some solutions for wavefront distortions canceling have already been proposed. In this paper we aim at improving an existing method for distorted curves retrieval, making use of a global optimization algorithm. We show that it is possible to combine an optimization method to an interpolation method in order to obtain a reliable and fast algorithm

The high resolution methods of array processing lead to an improvement of the results obtained for source localization. By adopting specific conventions, it is possible to employ high resolution methods to characterize straight lines in an image. In this paper, we propose an original method that leads to the estimation of the parameter "offset" of the straight lines. The proposed method is fast and effective compared to an existing method. An extension to non rectilinear contours is developed.