This paper proposes a novel robust Fisher Discriminant Analysis (FDA) for discriminative subspace learning in the presence of outliers. The proposed approach is motivated by the maximum-likelihood perspective on FDA and its connection to Kullback-Leibler (KL) divergence minimization. Within the probabilistic FDA framework, we develop a robust model by adopting the $\alpha $ -divergence as a flexible alternative to the KL divergence. The resulting method induces an adaptive redescending weighting scheme, in which each observation is weighted according to its statistical compatibility with the model, where the robustness level continuously controlled by $\alpha $ . As $\alpha $ decreases from 1, the influence of outliers is progressively suppressed, while classical FDA is recovered at $\alpha = 1$ . Combined with a two-fold iterative optimization procedure, the proposed method mitigates contamination at both the class-modeling stage and the projection-learning stage. We further provide theoretical analysis of the robustness mechanism, convergence, and computational complexity analysis to support the effectiveness and efficiency of the proposed method. Extensive experiments on synthetic data and multiple public image datasets under diverse contamination settings demonstrate that the proposed method consistently outperforms representative robust FDA variants and related approaches.

The dermatological concept of the Ugly-Duckling Sign (UDS) emphasizes the importance of comparing skin lesions within the same patient for enhanced diagnostic accuracy in melanoma detection, stating that atypical lesions are more likely to be cancers. However this concept is still underutilized in research, as most work on melanoma detection rely on classification ConvNets which lack the capacity to compare images together. Addressing this research gap, we introduce ReSeT (Residual Set-Transformer), a framework designed to compare skin lesions within patients during prediction. ReSeT comprises an encoder that takes individual images as input to generate embeddings, and a Set-Transformer with a residual prediction layer that compares these embeddings while predicting. We demonstrate that our architecture ReSeT significantly enhances performance compared to ConvNets and we highlighting the necessity of residual connections in the context of multi-output Transformers. We also observe that self-supervised encoders are able to generate embeddings of comparable quality to those of supervised models, showing their robustness and impact on image comparison tasks.

This paper presents our contribution to facial expression recognition using images obtained from the Light Field Face Dataset(LF). We compare several variants of neural network architectures to demonstrate the potential benefits of using this relatively new optical system in the field of facial expression recognition. We propose the use of the EfficientNetV2-S convolutional neural network as the base architecture, combined with various recurrent neural networks (LSTM, GRU, BiLSTM, and BiGRU) in our experiments. Furthermore, we investigate different sets of sub-aperture images, each varying in terms of the number of images and virtual position. The results demonstrate a significant improvement in accuracy for two specific configurations, depending on the sets of sub-aperture images used. The first configuration involves using the EfficientNetV2-S model in a two-branch configuration combined with an LSTM. The second configuration uses a single branch model with a BiLSTM.

Industry tends to optimize accuracy and time efficiency of every process analyzing its constraints and limits. Several tasks requiring high precision and reproducibility must be automated. In the context of secure Integrated Circuits (ICs) characterization, tasks such as power analysis are commonly automated. However, very few automations exist for tools calibration, while recent characterization schemes encounter mechanical constraints. Computer vision, flexible tool used in various fields, gives opportunities to address these constraints. In the case of laser fault injections, several positioning adjustments are required to ensure a maximal energy transmission in a targeted point. An accurate focalization of the laser beam is reached by using an autofocus system. Such a system is obtained by analyzing the camera view of the IC in the time-frequency domain. Whatever the method to disturb an IC, every secure characterization should be reproducible. By exploiting computer vision assistance, fault injection can be automated by mixing vision techniques to build a full or partial view of an IC and automatically identify the targeted IC and focus the perturbation on a chosen pattern in the image. A reliable pattern detection is implemented by studying spatial consistency of image's remarkable features represented by graphs. This paper presents several computer vision techniques to address above problems.

This paper addresses the detection of noisy structures in the context of infrared microscopy using labeled undirected graph matching. The selection of robust features as labels is determinant in this case of study, where hard conditions are dealt with: few relevant topological information, a potentially huge number of outliers and low contrasted images resulting in noisy graphs. Firstly, the image texture is reliably caught through a scale, rotation and intensity invariant histogram of oriented gradients. Secondly, in structures presenting numerous symmetries, the graph shape is locally registered while discriminated using weight from a super-increasing series. Coupled to the flexibility of the graph tensor product-based similarity metric, the matching framework achieves satisfying results.

Autofocus (AF) is a widely investigated subject in the fields of natural scene images, industrial assembly and biologic microscopy. This paper proposes a new effective AF method for infrared (IR) microscopy in the context of the Integrated Circuit industry (IC). The proposed method operates in the wavelet domain using a custom orthogonal wavelet for the 2D Discrete Wavelet Transform (DWT). The quality criterion of our AF algorithm relies on the standard deviance of the DWT coefficients, computed per subband and per level. Tested on several optical magnifying lenses, our method is robust time-efficient, and usable on-the-fly in the IC location system.

International audience

Autism spectrum disorders (ASD) is a disorder that affects communication, social skills or behaviours of some people. Children or adults with ASD often have some common repetitive behaviours or self-stimulatory behaviours. These behaviours usually refer to specific behaviours such as flapping, rocking, spinning, etc. This work investigates these behaviours and provides a benchmark dataset for researchers. In our knowledge, this dataset is the first 3D-dataset available online1in the area of 3D recognitions of complex and repetitive behaviours of autistic people. The 3D-Autism Dataset (3D-AD) is captured with Kinect sensor. We explore different categories of autistic repetitive behaviours: static and dynamic ones, simple and complex ones. Experiments have been done using dynamic time warping to detect these behaviours.

The purpose of this work is to propose a predictive model of a subject's age band. It is made under marketing constraints; these constraints define the customer's behavior towards five age bands. In this context, we first propose a regression approach, with Partial Least Squares Regression and Support Vector Regression. We then propose a classification approach by combining Support Vector Machines and Classification and Regression Tree. The classification approach overcomes the regression approach. In both approaches, we use Multi-Scale and Multilevel Histogram of Blocks of Local Binary Patterns (MMHLBP) as texture descriptor. Eventually, a comparison between our final model and the human operator shows the relevancy of the classification approach.

In this paper, we present a new person re-identification method based on appearance classification and silhouette part segmentation. In crowded areas, heads are considered as most apparent parts, hence the typical advantage of using the skeleton graph for the head detection and location of people after partial occlusion. The appearance classification consists in characterizing the appearance of a person into two classes, the frontal and the back appearance, using head detector and the orthogonal iteration algorithm for head pose estimation. The silhouette part segmentation divides the silhouette into three horizontal parts, ideally corresponding to head, torso and legs using skeleton graph and head detector. Our approach is robust to real world situations, in particular to variations in scales, human pose, illumination and clothes appearance changes. It also allows to reduce the confusion cases among people appearance and the amount of falsely matches.

In this paper, we present a new method for counting people. This method is based on the head detection after a segmentation of the human body by skeleton graph process. The skeleton silhouette is computed and decomposed into a set of segments corresponding to the head , torso and limbs. This structure captures the minimal information about the skeleton shape. No assumption is made about the viewpoint, this is done after the head pose process. Several results present the efficiency of the labelling process , particularly its structural properties for the detection of heads within a crowd. A proposed method are evaluated on the crowd counting task in the PETS 2010 dataset.

In this paper, we present a new method for counting people. This method is based on the head detection after a segmentation of the human body by skeleton graph process. The skeleton silhouette is computed and decomposed into a set of segments corresponding to the head, torso and limbs. This structure captures the minimal information about the skeleton shape. No assumption is made about the viewpoint, this is done after the head pose process. Several results present the efficiency of the labelling process , particularly its structural properties for the detection of heads within a crowd. A proposed method has been tested with an experiment of counting the number of pedestrians passing in a specific area.

Properly labeling human body parts in video sequences is essential for robust tracking and motion interpretation frameworks. We propose to perform this task by using Graph Matching. The silhouette skeleton is computed and decomposed into a set of segments corresponding to the different limbs. A Graph capturing the topology of the segments is generated and matched against a 3D model of the human skeleton. The limb identification is carried out for each node of the graph, potentially leading to the absence of correspondence. The method captures the minimal information about the skeleton shape. No assumption about the viewpoint, the human pose, the geometry or the appearance of the limbs is done during the matching process, making the approach applicable to every configuration. Some correspondences that might be ambiguous only relying on topology are enforced by tracking each graph node over time. Several results present the efficiency of the labeling, particularly its robustness to limb detection errors that are likely to occur in real situations because of occlusions or low level system failures. Finally the relevance of the labeling in an overall tracking system is described.

The described problem is a diagnostic task which must be solved for high-velocity industrial machine tools. Due to environmentally degraded conditions, direct measurements are not possible. Due to the rapidity of the machine, human intervention is not possible in case it causes positional faults. So, an oriented vision solution is proposed. The degraded conditions are vibrations, projections of water and metal chips, dazzling, etc. In this case, there is only one problem that is unable to achieve a diagnostic solution: is it the right piece in the right place? That is why complementary methods are proposed and presented in this paper to answer this question.