Edge-side Prognostics and Health Management (PHM) for Electric Vehicles (EVs) demands vehicle sub-systems Remaining Useful Life (RUL) predictors running on automotive-grade Microcontroller Units (MCUs) under tight memory, latency, and energy budgets. Classical Knowledge Distillation (KD) often degrades long-horizon accuracy and fails on the complex forecasting-to-decision pipeline for Silicon-Carbide (SiC) traction-inverter power modules. We propose Neuro-Modulated Knowledge Distillation (NM-KD), distilling a Multi-Scale Temporal Fusion Transformer (TFT-MS) into a lightweight model deployable on MCUs. NM-KD adapts loss weights via learned gates, bounds student sensitivity with a Lipschitz regularizer, and leverages Stochastic Weight Averaging (SWA) modulation. The TFT-MS teacher is benchmarked against several architectures spanning six design families, while the student selection is validated against five MCU-deployable alternatives. On SiC power devices, the student matches short-horizon performance and outperforms the teacher at long horizon (50-step accuracy 95.13% vs. 94.32%). On-target deployment reduces non-volatile memory by 83.12%, latency by 81.73%, and energy by 60.57%. Cross-batch generalization on 21 modules from two independent manufacturing lots confirms robustness to process variability. A lightweight temporal remapping converts accelerated-cycle predictions into field-relevant RUL estimates. NM-KD enables fast, low-power, accurate PHM on automotive Electronic Control Units (ECUs).

The automotive development of Electric Vehicles (EVs) is accelerating due to environmental concerns and technological advances. The complexity and performance demand of EVs require the use of Silicon-Carbide (SiC) technologies for the related efficiency, high-temperature resistance, and robust dynamic switching. To ensure the high performance of electric vehicles, the rigorous screening of the delivered SiC devices plays a pivotal role. Defects in SiC production can cause failures that directly affect the performance of electric vehicle engines, particularly in the traction inverter subsystem. To identify these defects early, advanced visual screening through optical microscopy and X-ray approaches has been proposed by leading car makers. These screening methodologies require advanced technical expertise and are still prone to significant errors, even when performed manually by experienced operators. To address these inefficiencies, the authors propose an ensemble deep learning pipeline for performing a robust automated visual inspection of SiC devices. By combining Multi-Head Attention blocks with adaptive input data distortion compensation through Jacobian regularization within convolutional architectures, the proposed model leverages global context and local feature maps, enhancing accuracy and robustness in multimodal defects detection. The proposed combined approach, tested on ACEPACK/TPACK DRIVE SiC power modules provided by STMicroelectronics, achieved an average accuracy of approximately 93% in both methodologies.

Continual learning has recently attracted attention from the research community, as it aims to solve long-standing limitations of classic supervised-trained models. However, most research on this subject has tackled continual learning in simple image classification scenarios. In this paper, we present a benchmark of state-of-the-art continual learning methods on video action recognition. Besides the increased complexity due to the temporal dimension, the video setting imposes stronger requirements on computing resources for top-performing rehearsal methods. To counteract the increased memory requirements, we present two method-agnostic variants for rehearsal methods, exploiting measures of either model confidence or data information to select memorable samples. Our experiments show that, as expected from the literature, rehearsal methods outperform other approaches; moreover, the proposed memory-efficient variants are shown to be effective at retaining a certain level of performance with a smaller buffer size.

Graph-structured scene descriptions can be efficiently used in generative models to control the composition of the generated image. Previous approaches are based on the combination of graph convolutional networks and adversarial methods for layout prediction and image generation, respectively. In this work, we show how employing multi-head attention to encode the graph information, as well as using a transformer-based model in the latent space for image generation can improve the quality of the sampled data, without the need to employ adversarial models with the subsequent advantage in terms of training stability. The proposed approach, specifically, is entirely based on transformer architectures both for encoding scene graphs into intermediate object layouts and for decoding these layouts into images, passing through a lower dimensional space learned by a vector-quantized variational autoencoder. Our approach shows an improved image quality with respect to state-of-the-art methods as well as a higher degree of diversity among multiple generations from the same scene graph. We evaluate our approach on three public datasets: Visual Genome, COCO, and CLEVR. We achieve an Inception Score of 13.7 and 12.8, and an FID of 52.3 and 60.3, on COCO and Visual Genome, respectively. We perform ablation studies on our contributions to assess the impact of each component. Code is available at https://github.com/perceivelab/trf-sg2im.

Polygonal meshes have become the standard for discretely approximating 3D shapes, thanks to their efficiency and high flexibility in capturing non-uniform shapes. This non-uniformity, however, leads to irregularity in the mesh structure, making tasks like segmentation of 3D meshes particularly challenging. Semantic segmentation of 3D mesh has been typically addressed through CNN-based approaches, leading to good accuracy. Recently, transformers have gained enough momentum both in NLP and computer vision fields, achieving performance at least on par with CNN models, supporting the long-sought architecture universalism. Following this trend, we propose a transformer-based method for semantic segmentation of 3D mesh motivated by a better modeling of the graph structure of meshes, by means of global attention mechanisms. In order to address the limitations of standard transformer architectures in modeling relative positions of non-sequential data, as in the case of 3D meshes, as well as in capturing the local context, we perform positional encoding by means the Laplacian eigenvectors of the adjacency matrix, replacing the traditional sinusoidal positional encodings, and by introducing clustering-based features into the self-attention and cross-attention operators. Experimental results, carried out on three sets of the Shape COSEG Dataset (Wang et al., 2012), on the human segmentation dataset proposed in Maron et al. (2017) and on the ShapeNet benchmark (Chang et al., 2015), show how the proposed approach yields state-of-the-art performance on semantic segmentation of 3D meshes.

Compared to other well-known biometric technologies based on physiological traits (e.g., fingerprint, iris, and face), heart biometrics are more robust to presentation attacks and are particularly suitable for continuous/periodic recognition.Most studies on heart biometrics concern electrocardiogram (ECG) and photoplethysmogram (PPG).While the reported results are encouraging, to the best of our knowledge, no studies have been conducted on the interoperability between ECG and PPG biometrics.We present a novel method that is capable of performing single-domain and multiple-domain identity verifications for ECG and PPG signals, providing interoperability between the heterogeneous cardiac signals.Our method does not require the computation of any reference/fiducial point and uses a compact representation of the given signals.We propose MultiCardioNet, a novel Siamese neural network trained by using an ad hoc learning algorithm.MultiCardioNet computes a similarity score between two spectrogram-based representations of cardiac signals.Our learning algorithm iteratively computes a balanced subset of genuine and impostor pairs during the training epochs.We performed experiments on a dataset containing 1,008 pairs of ECG and PPG samples, obtaining accuracy comparable to that of the state-of-the-art methods for single-domain scenarios and demonstrating only a relatively small performance decrease in the multiple-domain scenario.

Automotive industry is making rapid progress in the development of next generation cars with higher levels of autonomy and intelligent assistance. Although the general advanced driver assistance system (ADAS) architecture is widely discussed, limited interaction between driver and these intelligent solutions sometimes make these approaches inefficient. For these reasons, the authors triggered an investigation about driver's feedback in relation to the assistance inputs provided by the ADAS technologies. In this context, the goal of this proposal is the design of an intelligent system that learns from the analysis of the car driver eyes saccadic movements, the correlated level of attention towards the salient driving scene. With this approach, we enabled a visual-feedback system which learns the driver eye's fixing dynamic associated to the analyzed driving scene. Through ad-hoc enhanced motion magnification technique, the authors were able to amplify the mentioned saccadic dynamics in order to allow a downstream deep classifier to associate this physiological behavior with the corresponding level of the driver attention. The collected performances (over 97%) confirmed the effectiveness of the proposed method.

In the last few years, medical researchers have investigated promising approaches for cancer treatment, leading to a major interest in the immunotherapeutic approach. The target of immunotherapy is to boost a subject’s immune system in order to fight cancer. However, scientific studies confirmed that not all patients have a positive response to immunotherapy treatment. Medical research has long been engaged in the search for predictive immunotherapeutic-response biomarkers. Based on these considerations, we developed a non-invasive advanced pipeline with a downstream 3D deep classifier with attention and reinforcement learning for early prediction of patients responsive to immunotherapeutic treatment from related chest-abdomen CT-scan imaging. We have tested the proposed pipeline within a clinical trial that recruited patients with metastatic bladder cancer. Our experiment results achieved accuracy close to 93%.

In medical field, the term “immunotherapy” refers to a form of cancer treatment that uses the ability of body's immune system to prevent and destroy cancer cells. In the last few years, immunotherapy has demonstrated to be a very effective treatment in fighting cancer diseases. However, immunotherapy does not work for every patients and moreover, certain types of immunotherapy drugs could have side effects. With this regard, scientific researchers are investigating for effective ways to select the patients who are more likely to respond to the treatment. Hence, pre-clinical data confirmed that, sometimes, the composition of immune system cells infiltrating the tumor micro-environment may interfere with the efficacy of immunotherapy treatments. In this work, we developed a 3D Deep Network with a downstream classifier for selecting and properly augmenting features from chest-abdomen CT images toward improving cancer outcome prediction. In our work, we proposed an effective solution to a specific type of aggressive bladder cancer, called Metastatic Urothelial Carcinoma (mUC). Our experiment results achieved high accuracy confirming the effectiveness of the proposed pipeline.

Abstract In this work, we propose a 3D fully convolutional architecture for video saliency prediction that employs hierarchical supervision on intermediate maps (referred to as conspicuity maps) generated using features extracted at different abstraction levels. We provide the base hierarchical learning mechanism with two techniques for domain adaptation and domain-specific learning. For the former, we encourage the model to unsupervisedly learn hierarchical general features using gradient reversal at multiple scales, to enhance generalization capabilities on datasets for which no annotations are provided during training. As for domain specialization, we employ domain-specific operations (namely, priors, smoothing and batch normalization) by specializing the learned features on individual datasets in order to maximize performance. The results of our experiments show that the proposed model yields state-of-the-art accuracy on supervised saliency prediction. When the base hierarchical model is empowered with domain-specific modules, performance improves, outperforming state-of-the-art models on three out of five metrics on the DHF1K benchmark and reaching the second-best results on the other two. When, instead, we test it in an unsupervised domain adaptation setting, by enabling hierarchical gradient reversal layers, we obtain performance comparable to supervised state-of-the-art. Source code, trained models and example outputs are publicly available at https://github.com/perceivelab/hd2s .

Saliency detection is a fundamental process of human visual perception, since it allows us to identify the most important parts of a scene, directing our analysis and interpretation capabilities on a reduced set of information and reducing reaction times. However, current approaches for automatic saliency detection either attempt to mimic human capabilities by building attention maps from hand-crafted feature analysis, or employ convolutional neural networks trained as black boxes, without any architectural or information prior from human biology. In this paper, we present an approach for saliency detection that combines the success of deep learning in identifying representations for visual data with a training paradigm aimed at matching neural activity provided directly by brain signals recorded while subjects look at images. We show that our approach is able to capture correspondences between visual elements and neural activities, successfully generalizing to unseen images to identify their most salient regions.

In this paper we define a deep learning architecture for automated segmentation of anatomical structures in Craniomaxillofacial (CMF) CT scans that leverages the recent success of encoder-decoder models for semantic segmentation of natural images. In particular, we propose a fully convolutional deep network that combines the advantages of recent fully convolutional models, such as Tiramisu, with squeeze-and-excitation blocks for feature recalibration, integrated with convolutional LSTMs to model spatio-temporal correlations between consecutive slices. The proposed segmentation network shows superior performance and generalization capabilities (to different structures and imaging modalities) than state of the art methods on automated segmentation of CMF structures (e.g., mandibles and airways) in several standard benchmarks (e.g., MICCAI datasets) and on new datasets proposed herein, effectively facing shape variability.

In this paper we define a deep multi-stage architecture for automated landmarking of craniomaxillofacial (CMF) CT images. Our model is composed of three subnetworks that first localize, on reduced-resolution images, areas where landmarks may be found and then refine the search, at full-resolution scale, through a hierarchical structure aiming at increasing the granularity of the investigated region. The multi-stage pipeline is designed to deal with full resolution data and does not require any additional pre-processing step to reduce search space, as opposed to existing methods that can be only adopted for searching landmarks located in well-defined anatomical structures (e.g., mandibles). The automated landmarking system is tested on identifying landmarks located in several CMF regions, achieving an average error of 0.8 mm, significantly lower than expert readings. The proposed model also outperforms baselines and is on par with existing models that employ additional upstream segmentation, on state-of-the-art benchmarks.

Being able to estimate the traversability of the area surrounding a mobile robot is a fundamental task in the design of a navigation algorithm. However, the task is often complex, since it requires evaluating distances from obstacles, type and slope of terrain, and dealing with non-obvious discontinuities in detected distances due to perspective. In this paper, we present an approach based on deep learning to estimate and anticipate the traversing score of different routes in the field of view of an on-board RGB camera. The backbone of the proposed model is based on a state-of-the-art deep segmentation model, which is fine-tuned on the task of predicting route traversability. We then enhance the model's capabilities by a) addressing domain shifts through gradient-reversal unsupervised adaptation, and b) accounting for the specific safety requirements of a mobile robot, by encouraging the model to err on the safe side, i.e., penalizing errors that would cause collisions with obstacles more than those that would cause the robot to stop in advance. Experimental results show that our approach is able to satisfactorily identify traversable areas and to generalize to unseen locations.

Integrating human-provided location priors into video object segmentation has been shown to be an effective strategy to enhance performance, but their application at large scale is unfeasible. Gamification can help reduce the annotation burden, but it still requires user involvement. We propose a video object segmentation framework that leverages the combined advantages of user feedback for segmentation and gamification strategy by simulating multiple game players through a reinforcement learning (RL) model that reproduces human ability to pinpoint moving objects and using the simulated feedback to drive the decisions of a fully convolutional deep segmentation network. Experimental results on the DAVIS-17 benchmark show that: 1) including user-provided prior, even if not precise, yields high performance; 2) our RL agent replicates satisfactorily the same variability of humans in identifying spatiotemporal salient objects; and 3) employing artificially generated priors in an unsupervised video object segmentation model reaches state-of-the-art performance.

Recently, the ability to monitor driver drowsiness has attracted a great deal of attention in the automotive industry, in order to prevent the risk due to an inadequate driver psycho-physical state. Specifically, the research effort has focused on the study of the physiological signals to assess the attention level. The main idea consists in verifying the drowsiness level through analyzing the Heart Rate Variability (HRV). The HRV allows to understand the activity of the autonomic nervous system that regulates a series of unconscious and involuntary activities (e.g. the heartbeat, the blood pressure). The HRV is traditionally obtained from electrocardiography (ECG) even though the photoplethysmography (PPG) signal has been proposed as valid alternative to ECG in order to overcome some limitations derived from it. For the above reasons, we analyzed the skin micro-movements and changes in facial color due to blood circulation quite indistinguishable with naked eye in order to extract facial landmarks and to reconstruct PPG signal. The results we obtained by validation confirmed the correlation between the PPG signal detected by sensors and the reconstructed PPG signal from facial landmarks.

Traditional methods for early detection of melanoma rely on the visual analysis of the skin lesions performed by a dermatologist. The analysis is based on the so‐called ABCDE (Asymmetry, Border irregularity, Colour variegation, Diameter, Evolution) criteria, although confirmation is obtained through biopsy performed by a pathologist. The proposed method exploits an automatic pipeline based on morphological analysis and evaluation of skin lesion dermoscopy images. Preliminary segmentation and pre‐processing of dermoscopy image by SC‐cellular neural networks is performed, in order to obtain ad‐hoc grey‐level skin lesion image that is further exploited to extract analytic innovative hand‐crafted image features for oncological risks assessment. In the end, a pre‐trained Levenberg–Marquardt neural network is used to perform ad‐hoc clustering of such features in order to achieve an efficient nevus discrimination (benign against melanoma), as well as a numerical array to be used for follow‐up rate definition and assessment. Moreover, the authors further evaluated a combination of stacked autoencoders in lieu of the Levenberg–Marquardt neural network for the clustering step.

The DNA microarray images allow to analyze the natural gene expressions. In this paper we propose an advanced method to efficiently address the imaging storage as well as the performance of the algorithm used to retrieve information from DNA images. The cellular neural networks (CNNs) based core is able to provide a method to extract foreground (the DNA gene expression information) from DNA images. It is also proposed an innovative method to compress the DNA image by re-organizing the signal data belonging to the background by making use of a novel way to apply the re-indexing techniques to almost ¿uncorrelated¿ signal. Experiments confirm how the proposed method outperform previous solution in almost all cases.

Palette re-ordering is a well known and very effective approach for improving the compression of color indexed images. If the spatial distribution of the indexes in the image is smooth, greater compression ratios may be obtained. As known, obtaining an optimal re-indexing scheme is not a trivial task. In this paper we provide a novel algorithm for palette re-ordering problem making use of a motor map neural network. Experimental results show the real effectiveness of the proposed method both in terms of compression ratio and zero-order entropy of local differences. Also its computational complexity is competitive with previous works in the field.

Palette re-ordering is an effective approach for improving the compression of color-indexed images. If the spatial distribution of the indexes in the image is smooth, greater compression ratios may be obtained. As is already known, obtaining an optimal re-indexing scheme is not a trivial task. In this paper, we provide a novel algorithm for palette re-ordering problem making use of a motor map neural network. Experimental results show the real effectiveness of the proposed method both in terms of compression ratio and zero-order entropy of local differences. Also, its computational complexity is competitive with previous works in the field.