Super-Resolution (SISR) is a computer vision task that aims to generate high-resolution images from their low-resolution counterparts. Typically, Super-Resolution methods use scaling factors of x2, x3, or x4 to uniformly enhance the resolution of the entire image. However, some acquisition devices, such as 360 ^∘ cameras, produce images with non-uniform resolution across the frame. In this work, we propose to adapt state-of-the-art efficient methods for Single Image Super-Resolution to address the challenge of restoring images affected by spatially varying degradations. Specifically, we focus on the method that won the recent NTIRE 2024 Efficient Super-Resolution Challenge. For our experiments, synthetic images with different spatially varying types of degradation are generated, and the SISR method is specifically modified and trained to effectively handle such challenging scenarios. In addition to evaluating the developed method with traditional image quality metrics such as PSNR and SSIM, we also assess its practical impact on a downstream object detection task. The results on the WIDER FACE face-detection dataset, using the YOLOv8 object detection model, show that applying the proposed SISR approach to images with spatially varying degradations produces artifact-free outputs and enables object detectors to achieve superior performance compared to their application on degraded images.

The widespread sharing of face images on social media platforms and in large-scale datasets raises pressing privacy concerns, as biometric identifiers can be exploited without consent. Face anonymization seeks to generate realistic facial images that irreversibly conceal the subject's identity while preserving their usefulness for downstream tasks. However, most existing generative approaches focus on identity removal and image realism, often neglecting facial expressions as well as photometric consistency – specifically attributes such as illumination and skin tone – that are critical for applications like relighting, color constancy, and medical or affective analysis. In this work, we propose a feature-preserving anonymization framework that extends DeepPrivacy by incorporating dense facial landmarks to better retain expressions, and by introducing lightweight post-processing modules that ensure consistency in lighting direction and skin color. We further establish evaluation metrics specifically designed to quantify expression fidelity, lighting consistency, and color preservation, complementing standard measures of image realism, pose accuracy, and re-identification resistance. Experiments on the CelebA-HQ dataset demonstrate that our method produces anonymized faces with improved realism and significantly higher fidelity in expression, illumination, and skin tone compared to state-of-the-art baselines. These results underscore the importance of feature-aware anonymization as a step toward more useful, fair, and trustworthy privacy-preserving facial data.

We present FLIC, a real-world annotated dataset designed for the visual estimation of food leftovers in canteens and other collective catering environments using standard 2D RGB imagery. Collected over 22 days in an operational university canteen, the dataset includes 401 paired image acquisitions of full and leftover trays, each associated with pixel-precise semantic segmentation masks and physically measured food mass. The goal is to support research on the estimation of leftover food mass from tray images, a task that has received limited attention compared to pre-consumption food recognition, despite its relevance for sustainability and operational decision making in food services. Unlike existing food datasets, FLIC jointly provides paired before–after visual observations and reliable mass ground truth, enabling quantitative analysis of food leftovers under realistic conditions without relying on depth or multi-view information. To demonstrate the dataset’s applicability, we rely on the concept of digital density, relating pixel area to food mass, and implement a lightweight, interpretable baseline mass estimation pipeline. This includes an automatic food/no-food segmentation stage, evaluated across multiple deep learning models (U-Net, DABNet, DINOv2+FeatUp, and SAM), followed by an assisted food recognition stage that leverages the fixed daily menu to map broad user input (e.g., “first course” vs. “second course”) to a specific food class. Experimental results highlight both the potential and the intrinsic challenges of visual food leftover estimation.

Recent advances in snapshot multispectral (MS) imaging have enabled compact, low-cost spectral sensors for consumer and mobile devices. By capturing richer spectral information than conventional RGB sensors, these systems can enhance key imaging tasks, including color correction. However, most existing methods treat the color correction pipeline in separate stages, often discarding MS data early in the process. We propose a unified, learning-based framework that (i) performs end-to-end color correction and (ii) jointly leverages data from a high-resolution RGB sensor and an auxiliary low-resolution MS sensor. Our approach integrates the full pipeline within a single model, producing coherent and color-accurate outputs. We demonstrate the flexibility and generality of our framework by refactoring two different state-of-the-art image-to-image architectures. To support training and evaluation, we construct a dedicated dataset by aggregating and repurposing publicly available spectral datasets, rendering under multiple RGB camera sensitivities. Extensive experiments show that our approach improves color accuracy and stability, reducing error by up to 50\% compared to RGB-only and MS-driven baselines. Datasets, code, and models will be made available upon acceptance.

Computational color constancy (CC) aims to recover the true colors of a scene despite changes in illumination. While recent cross-camera deep learning methods have made impressive progress, their training relies on relatively small RAW datasets. In this work, we introduce a zero-shot color constancy framework grounded in intrinsic image decomposition (IID). Our method estimates the albedo and infers the illumination map from a pre-trained IID model, followed by chromaticity clustering. This design allows us to avoid any fine-tuning on CC datasets. Experiments on the Gehler-Shi and NUS-8 datasets show that our approach achieves performance comparable to cross-dataset SOTA performance, while maintaining robustness to camera and dataset variations. These results highlight the potential of transferring IID-learned illumination priors to achieve truly sensor-agnostic color constancy.

Hyperspectral image super-resolution (HSI-SR) aims to reconstruct hyperspectral images at high spatial resolution, starting from low-resolution inputs, while preserving both spatial details and spectral fidelity. In this work, we propose the Efficient Spatial-Spectral Processing Network (ESSPN), a lightweight deep learning architecture designed to address the challenges of HSI-SR in a computationally efficient way. ESSPN is built around a novel Spatial-Spectral Block (SSB) that separately models spatial structures and spectral correlations through residual convolutional and attention mechanisms. The network head incorporates an efficient upsampling module based on pixel shuffle decomposition to produce high-resolution outputs without interpolation artifacts. Extensive experiments on two publicly available datasets, i.e., ARAD1K and StereoMSI demonstrate that ESSPN achieves competitive or superior performance compared to state-of-the-art methods when evaluated at scale factors of × 4, × 6 and × 8. Notably, the model shows strong generalization across hyperspectral cameras with varying spectral responses, and across radiometric domains, covering both radiance and reflectance measurements, while requiring significantly fewer parameters and FLOPs compared to existing methods. These results position the proposed ESSPN as a practical and effective solution for high-quality hyperspectral image super-resolution in real-world applications.

Hyperspectral imaging (HSI) enables fine spectral analysis but is often limited by low spatial resolution due to sensor constraints. To address this, we propose CGNet, a color-guided hyperspectral super-resolution network that leverages complementary information from low-resolution hyperspectral inputs and high-resolution RGB images. CGNet adopts a dual-encoder design: the RGB encoder extracts hierarchical spatial features, while the HSI encoder progressively upsamples spectral features. A multi-scale fusion decoder then combines both modalities in a coarse-to-fine manner to reconstruct the high-resolution HSI. Training is driven by a hybrid loss that balances L1 and Spectral Angle Mapper (SAM), which ablation studies confirm as the most effective formulation. Experiments on two benchmarks, ARAD1K and StereoMSI, at ×4 and ×6 upscaling factors demonstrate that CGNet consistently outperforms state-of-the-art baselines. CGNet achieves higher PSNR and SSIM, lower SAM, and reduced ΔE00, confirming its ability to recover sharp spatial structures while preserving spectral fidelity.

The classification of distracted drivers is pivotal for ensuring safe driving. Previous studies demonstrated the effectiveness of neural networks in automatically predicting driver distraction, fatigue, and potential hazards. However, recent research has uncovered a significant loss of accuracy in these models when applied to samples acquired under conditions that differ from the training data. In this paper, we introduce a robust model designed to withstand changes in camera position within the vehicle. Our Driver Behavior Monitoring Network (DBMNet) relies on a lightweight backbone and integrates a disentanglement module to discard camera view information from features, coupled with contrastive learning to enhance the encoding of various driver actions. Experiments conducted using a leave-one-camera-out protocol on the daytime and nighttime subsets of the 100-Driver dataset validate the effectiveness of our approach. Cross-dataset and cross-camera experiments conducted on three benchmark datasets, namely AUCDD-V1, EZZ2021 and SFD, demonstrate the superior generalization capabilities of the proposed method. Overall DBMNet achieves an improvement of 7% in Top-1 accuracy compared to existing efficient approaches. Moreover, a quantized version of the DBMNet and all considered methods has been deployed on a Coral Dev Board board. In this deployment scenario, DBMNet outperforms alternatives, achieving the lowest average error while maintaining a compact model size, low memory footprint, fast inference time, and minimal power consumption.

Early and accurate diagnosis of Alzheimer’s Disease (AD) is crucial for patient care and intervention but remains challenging due to the complexity and variability of clinical data. Electroencephalography (EEG) has emerged as a promising, non-invasive, and cost-effective tool to detect brain activity patterns associated with AD. In this work, we introduce a lightweight Multiscale Temporal Deep Network (MTDNet) that integrates multiple temporal convolutions with recurrent modeling to capture both short- and long-term EEG patterns. Two patient-level classification strategies are also proposed that combine segment-level EEG predictions based on consensus and score aggregation to better align with clinical practice and utility. We evaluate our method on four benchmark EEG datasets (ADSZ, APAVA, ADFTD, BrainLat) where it consistently outperforms state-of-the-art solutions by about 2% at the segment level and by about 6% at the patient level on the most challenging datasets. Unlike recent computationally heavy transformer-based solutions, MTDNet achieves superior accuracy with only 20.5K parameters and 1.8M FLOPs, enabling its deployment in resource-constrained environments. Ablation studies confirm the critical contribution of the multiscale design and show that simple augmentation techniques increase generalization and robustness. Code is available at https://github.com/unimib-islab/MTDNet.

Self-supervised learning has recently gained increasing attention in computer vision, enabling the extraction of rich and general-purpose feature representations without requiring large annotated datasets. In this paper we aim to build a unified approach capable of deploying robust and effective analysis systems, replacing the need for multiple task-specific models trained end-to-end. Rather than introducing new architectures or training strategies, our goal is to systematically assess whether a single frozen self-supervised representation can support heterogeneous food-related tasks under realistic operating conditions. To this end, we performed an extensive analysis of DINOv2 features across multiple benchmark datasets and tasks, including food classification, segmentation, aesthetic assessment, and robustness to image distortions. In addition, we explore its capacity for continual learning by applying it to incremental food classification scenarios. Our findings reveal that DINOv2 features excel in many food-related applications. Their shared representations across tasks reduce the need for training separate models, while their strong generalization, high accuracy, and ability to handle complex multi-task scenarios make them a strong candidate for a unified food recognition approach. Specifically, DINOv2 features match or surpass state-of-the-art supervised methods in several food recognition tasks, while offering a simpler and more unified deployment strategy. Furthermore, they outperform end-to-end models in cross-dataset scenarios by up to +19.4% Top-1 accuracy and exhibits strong resilience to common image distortions by up to +48.0% robustness in Top-1 accuracy percentual difference, ensuring reliable performance in real-world applications. On average across all considered tasks, the DINOv2-based unified evaluation outperforms the state of the art by approximately 2.8% and 5.4%, depending on the chosen model size, while using only 6.2% and 23.9% of the total number of model parameters, respectively.

This paper introduces CLOE, a confidence-based local-to-global fully convolutional framework for multispectral illuminant estimation, which jointly captures spatial structure and spectral correlations through adaptive confidence-weighted pooling. The architecture integrates a Spatial Feature Extractor and a Spectral Branch through a confidence-based fusion mechanism that produces robust global illuminant predictions. Evaluations on the KAUST and BeyondRGB datasets demonstrate that CLOE consistently outperforms both classical statistical methods and recent learning-based approaches, achieving state-of-the-art performance in terms of angular error between the estimated and ground-truth illuminant spectra (mean-∆AMS and std-∆AMS). Qualitative analyses further show that CLOE provides interpretable intermediate representations, enabling visual inspection of spatial–spectral cues. These results highlight the effectiveness and robustness of the proposed method for real-world multispectral imaging.

Underwater object detection in side-scan sonar (SSS) imagery is a challenging task due to limited labeled data and the spatially varying resolution inherent to SSS images. To address these issues, we generate a large-scale synthetic dataset of labeled SSS images using Blender, designed for training and evaluating mine detection models. With a total of 10,000 images, it exceeds existing public datasets by one order of magnitude. The dataset is used to investigate whether applying single image super-resolution (SISR) techniques, such as SPAN and DRCT, to create input images with uniform spatial resolution can enhance detection performance; moreover, SISR can also be used to standardize inputs across data sources, thus reducing the need to retrain detectors for new sensors. Experiments are conducted using the YOLOv11 detector, comparing detection results on both native synthetic images and their super-resolved counterparts. The experimental results show that super-resolution pre-processing permits to achieve detection performance comparable to, and occasionally surpassing, that on high-resolution synthetic data. These findings demonstrate the utility of synthetic datasets in overcoming data scarcity and highlight the potential of SISR in improving underwater object detection in scenarios with spatially varying image quality.

Color correction is a key component of camera image signal processing (ISP) pipelines, encompassing illuminant discounting and colorimetric mapping of device-dependent sensor responses to device-independent color spaces, such as CIE XYZ. Despite extensive research, accurate color correction remains challenging due to the non-linear relationship between camera sensor responses and CIE XYZ color space, as well as to the increasing presence of highly chromatic and spectrally complex LED illuminants. We propose a color correction framework based on illuminant-adaptive three-dimensional lookup tables (LUTs), which we call Color Correction LUT (C^2LUT). Our method combines a chromaticity-aware illuminant representation with a non-linear color transformation, enabling accurate correction under illuminants spanning a wide range of chromaticities and spectral complexities. We employ Tucker tensor decomposition to represent the LUTs, ensuring that computational requirements remain sufficiently low for deployment in camera ISPs. In addition, we introduce a large-scale illuminants dataset comprising 1,473 spectral power distributions, with different chromaticities and spectral profiles. Experiments across multiple cameras, illuminants, reflectance datasets, and real captured images demonstrate consistent improvements over existing methods for color correction, reducing CIE ΔE_00 by up to 20

Traditional Image Quality Assessment (IQA) has primarily aimed to quantify perceptual quality in terms of technical degradations such as noise, blur, or compression artifacts. However, in image rendering, the key factor influencing perceived quality is not the presence of degradations but the manner in which color processing algorithms are applied, as they directly shape the overall aesthetic appearance of the image. To date, the quantitative evaluation of how rendering methods affect image quality has been insufficiently addressed. In this work, we introduce Image Rendering Quality Assessment (IRQA) as a new problem setting within IQA and present REPID, a benchmark designed for its study. REPID contains 30,000 edited images and preference annotations collected from 13,648 voters, resulting in an over 2.5 million unique votes. Based on REPID, we investigate content-dependent render preferences and the influence of rendering parameters, and further explore applications such as aesthetic preference prediction (including personalization), render ranking, and benchmarking of aesthetic evaluation methods. We further perform an extensive experimental comparison of traditional IQA metrics, handcrafted features, deep learning approaches, and foundation-model embeddings. On the REPID benchmark, IRQA-specific models achieve up to 40% better precision than conventional distortion-oriented IQA methods.

Understanding driving behavior is critical for enhancing road safety, optimizing insurance, healthcare costs, and informing the development of mobility services and advanced vehicle systems. Driving remains a high-risk daily activity, influenced by a complex interplay of psychological, cognitive, and contextual factors such as stress, impulsivity, risk perception, and environmental conditions. To address this complexity, this research empirically investigated a broad range of psychological and behavioral aspects of driving through a combined methodology involving standardized psychometric surveys and high-fidelity driving simulation. The experimental setup integrates wearable biometric sensors to monitor psycho-physiological responses under varying levels of cognitive load and environmental complexity. Findings will inform the development of cognitively enriched software agents, enabling more realistic agent-based traffic simulations capturing both vehicle dynamics and the psychological dimensions of driving behavior. Preliminary results, mainly aimed to verify the correctness of the experimental design, are presented and discussed to assess the capability of reproducing nuanced driving behaviors within agent-based traffic simulations to enhance their ecological validity, adaptability, and predictive accuracy.

Multimodal Large Language Models (MLLMs) have achieved notable performance in computer vision tasks that require reasoning across visual and textual modalities, yet their capabilities are limited to their pre-trained data, requiring extensive fine-tuning for updates. Recent researches have explored the use of In-Context Learning (ICL) to overcome these challenges by providing a set of demonstrating examples as context to augment MLLMs performance in several tasks, showing that many-shot ICL leads to substantial improvements compared to few-shot ICL. However, the reliance on numerous demonstrating examples and the limited MLLMs context windows presents significant obstacles. This paper aims to address these challenges by introducing a novel approach, Visual RAG, that synergically combines the MLLMs capability to learn from the context, with a retrieval mechanism. The crux of this approach is to ensure to augment the MLLM knowledge by selecting only the most relevant demonstrating examples for the query, pushing it to learn by analogy. In this way, relying on the new information provided dynamically during inference time, the resulting system is not limited to the knowledge extracted from the training data, but can be updated rapidly and easily without fine-tuning. Furthermore, this greatly reduces the computational costs for improving the model image classification performance, and augments the model knowledge to new visual domains and tasks it was not trained for. Extensive experiments on eight different datasets in the state of the art spanning several domains and image classification tasks show that the proposed Visual RAG, compared to the most recent state of the art (i.e., many-shot ICL), is able to obtain an accuracy that is very close or even higher (approx. +2 using a much smaller set of demonstrating examples (approx. only 23 average).

Deep learning has revolutionized computer vision by allowing neural networks to automatically learn features from data. However, the highly nonlinear nature of deep neural networks makes them difficult to interpret, leading to concerns about potential biases in critical applications. To address this, researchers have advocated for eXplainable Artificial Intelligence (XAI). Many XAI techniques have been proposed but all of them only highlight image regions influencing model decisions, lacking any further explanations. In this paper, we propose a post-hoc model-agnostic meta-XAI method that explains why specific image regions are used for decisions. The paper presents the experimental setup and results, discussing the perturbations used for explanations in color, frequency, shape, shading, and texture. The explanation is given in terms of human-interpretable image features, e.g., color, shape, shading, and texture both as perturbation plots and as visual summary through the use of the newly introduced normalized Area Under the Curve score. The experimental results confirm the previous findings that vision deep learning models are biased towards texture, but also highlight the importance of color, frequency content and perceptually salient structures in the final decision.

Traditional Image Quality Assessment (IQA) focuses on quantifying technical degradations such as noise, blur, or compression artifacts, using both full-reference and no-reference objective metrics. However, evaluation of rendering aesthetics, a growing domain relevant to photographic editing, content creation, and AI-generated imagery, remains underexplored due to the lack of datasets that reflect the inherently subjective nature of style preference. In this work, a novel benchmark dataset designed to model human aesthetic judgments of image rendering styles is introduced: the Dataset for Evaluating the Aesthetics of Rendering (DEAR). Built upon the MIT-Adobe FiveK dataset, DEAR incorporates pairwise human preference scores collected via large-scale crowdsourcing, with each image pair evaluated by 25 distinct human evaluators with a total of 13,648 of them participating overall. These annotations capture nuanced, context-sensitive aesthetic preferences, enabling the development and evaluation of models that go beyond traditional distortion-based IQA, focusing on a new task: Evaluation of Aesthetics of Rendering (EAR). The data collection pipeline is described, human voting patterns are analyzed, and multiple use cases are outlined, including style preference prediction, aesthetic benchmarking, and personalized aesthetic modeling. To the best of the authors' knowledge, DEAR is the first dataset to systematically address image aesthetics of rendering assessment grounded in subjective human preferences. A subset of 100 images with markup for them is published on HuggingFace (huggingface.co/datasets/vsevolodpl/DEAR).

This paper presents a comprehensive review of the NTIRE 2025 Challenge on Single-Image Efficient Super-Resolution (ESR). The challenge aimed to advance the development of deep models that optimize key computational metrics, i.e., runtime, parameters, and FLOPs, while achieving a PSNR of at least 26.90 dB on the DIV2K_LSDIR_valid dataset and 26.99 dB on the DIV2K_LSDIR_test dataset. A robust participation saw 244 registered entrants, with 43 teams submitting valid entries. This report meticulously analyzes these methods and results, emphasizing groundbreaking advancements in state-of-the-art single-image ESR techniques. The analysis highlights innovative approaches and establishes benchmarks for future research in the field.