Jean-Michel Morel

Bruno Galerne, Lara Raad, José Lezama et al.

Neural style transfer (NST) is a deep learning technique that produces an unprecedentedly rich style transfer from a style image to a content image. It is particularly impressive when it comes to transferring style from a painting to an image. NST was originally achieved by solving an optimization problem to match the global statistics of the style image while preserving the local geometric features of the content image. The two main drawbacks of this original approach is that it is computationally expensive and that the resolution of the output images is limited by high GPU memory requirements. Many solutions have been proposed to both accelerate NST and produce images with larger size. However, our investigation shows that these accelerated methods all compromise the quality of the produced images in the context of painting style transfer. Indeed, transferring the style of a painting is a complex task involving features at different scales, from the color palette and compositional style to the fine brushstrokes and texture of the canvas. This paper provides a solution to solve the original global optimization for ultra-high resolution (UHR) images, enabling multiscale NST at unprecedented image sizes. This is achieved by spatially localizing the computation of each forward and backward passes through the VGG network. Extensive qualitative and quantitative comparisons, as well as a \textcolor{coverletter}{perceptual study}, show that our method produces style transfer of unmatched quality for such high-resolution painting styles. By a careful comparison, we show that state-of-the-art fast methods are still prone to artifacts, thus suggesting that fast painting style transfer remains an open problem. Source code is available at https://github.com/bgalerne/scaling_painting_style_transfer.

Thomas Eboli, Jean-Michel Morel, Gabriele Facciolo

The optics of any camera degrades the sharpness of photographs, which is a key visual quality criterion. This degradation is characterized by the point-spread function (PSF), which depends on the wavelengths of light and is variable across the imaging field. In this paper, we propose a two-step scheme to correct optical aberrations in a single raw or JPEG image, i.e., without any prior information on the camera or lens. First, we estimate local Gaussian blur kernels for overlapping patches and sharpen them with a non-blind deblurring technique. Based on the measurements of the PSFs of dozens of lenses, these blur kernels are modeled as RGB Gaussians defined by seven parameters. Second, we remove the remaining lateral chromatic aberrations (not contemplated in the first step) with a convolutional neural network, trained to minimize the red/green and blue/green residual images. Experiments on both synthetic and real images show that the combination of these two stages yields a fast state-of-the-art blind optical aberration compensation technique that competes with commercial non-blind algorithms.

Adrien Courtois, Jean-Michel Morel, Pablo Arias

Neural networks trained on large datasets by minimizing a loss have become the state-of-the-art approach for resolving data science problems, particularly in computer vision, image processing and natural language processing. In spite of their striking results, our theoretical understanding about how neural networks operate is limited. In particular, what are the interpolation capabilities of trained neural networks? In this paper we discuss a theorem of Domingos stating that "every machine learned by continuous gradient descent is approximately a kernel machine". According to Domingos, this fact leads to conclude that all machines trained on data are mere kernel machines. We first extend Domingo's result in the discrete case and to networks with vector-valued output. We then study its relevance and significance on simple examples. We find that in simple cases, the "neural tangent kernel" arising in Domingos' theorem does provide understanding of the networks' predictions. Furthermore, when the task given to the network grows in complexity, the interpolation capability of the network can be effectively explained by Domingos' theorem, and therefore is limited. We illustrate this fact on a classic perception theory problem: recovering a shape from its boundary.

Adrien Courtois, Jean-Michel Morel, Pablo Arias

Recent years have seen the emergence of many new neural network structures (architectures and layers). To solve a given task, a network requires a certain set of abilities reflected in its structure. The required abilities depend on each task. There is so far no systematic study of the real capacities of the proposed neural structures. The question of what each structure can and cannot achieve is only partially answered by its performance on common benchmarks. Indeed, natural data contain complex unknown statistical cues. It is therefore impossible to know what cues a given neural structure is taking advantage of in such data. In this work, we sketch a methodology to measure the effect of each structure on a network's ability, by designing ad hoc synthetic datasets. Each dataset is tailored to assess a given ability and is reduced to its simplest form: each input contains exactly the amount of information needed to solve the task. We illustrate our methodology by building three datasets to evaluate each of the three following network properties: a) the ability to link local cues to distant inferences, b) the translation covariance and c) the ability to group pixels with the same characteristics and share information among them. Using a first simplified depth estimation dataset, we pinpoint a serious nonlocal deficit of the U-Net. We then evaluate how to resolve this limitation by embedding its structure with nonlocal layers, which allow computing complex features with long-range dependencies. Using a second dataset, we compare different positional encoding methods and use the results to further improve the U-Net on the depth estimation task. The third introduced dataset serves to demonstrate the need for self-attention-like mechanisms for resolving more realistic depth estimation tasks.

Yu Guo, Qiyu Jin, Jean-Michel Morel et al.

Image demosaicing and denoising play a critical role in the raw imaging pipeline. These processes have often been treated as independent, without considering their interactions. Indeed, most classic denoising methods handle noisy RGB images, not raw images. Conversely, most demosaicing methods address the demosaicing of noise free images. The real problem is to jointly denoise and demosaic noisy raw images. But the question of how to proceed is still not yet clarified. In this paper, we carry-out extensive experiments and a mathematical analysis to tackle this problem by low complexity algorithms. Indeed, both problems have been only addressed jointly by end-to-end heavy weight convolutional neural networks (CNNs), which are currently incompatible with low power portable imaging devices and remain by nature domain (or device) dependent. Our study leads us to conclude that, with moderate noise, demosaicing should be applied first, followed by denoising. This requires a simple adaptation of classic denoising algorithms to demosaiced noise, which we justify and specify. Although our main conclusion is ``demosaic first, then denoise'', we also discover that for high noise, there is a moderate PSNR gain by a more complex strategy: partial CFA denoising followed by demosaicing, and by a second denoising on the RGB image. These surprising results are obtained by a black-box optimization of the pipeline, which could be applied to any other pipeline. We validate our results on simulated and real noisy CFA images obtained from several benchmarks.

Valéry Dewil, Zhe Zheng, Arnaud Barral et al.

MIMO (multiple input, multiple output) approaches are a recent trend in neural network architectures for video restoration problems, where each network evaluation produces multiple output frames. The video is split into non-overlapping stacks of frames that are processed independently, resulting in a very appealing trade-off between output quality and computational cost. In this work we focus on the low-latency setting by limiting the number of available future frames. We find that MIMO architectures suffer from problems that have received little attention so far, namely (1) the performance drops significantly due to the reduced temporal receptive field, particularly for frames at the borders of the stack, (2) there are strong temporal discontinuities at stack transitions which induce a step-wise motion artifact. We propose two simple solutions to alleviate these problems: recurrence across MIMO stacks to boost the output quality by implicitly increasing the temporal receptive field, and overlapping of the output stacks to smooth the temporal discontinuity at stack transitions. These modifications can be applied to any MIMO architecture. We test them on three state-of-the-art video denoising networks with different computational cost. The proposed contributions result in a new state-of-the-art for low-latency networks, both in terms of reconstruction error and temporal consistency. As an additional contribution, we introduce a new benchmark consisting of drone footage that highlights temporal consistency issues that are not apparent in the standard benchmarks.

Franco Marchesoni-Acland, Jean-Michel Morel, Josselin Kherroubi et al.

Even though data annotation is extremely important for interpretability, research and development of artificial intelligence solutions, most research efforts such as active learning or few-shot learning focus on the sample efficiency problem. This paper studies the neglected complementary problem of getting annotated data given a predictor. For the simple binary classification setting, we present the spectrum ranging from optimal general solutions to practical efficient methods. The problem is framed as the full annotation of a binary classification dataset with the minimal number of yes/no questions when a predictor is available. For the case of general binary questions the solution is found in coding theory, where the optimal questioning strategy is given by the Huffman encoding of the possible labelings. However, this approach is computationally intractable even for small dataset sizes. We propose an alternative practical solution based on several heuristics and lookahead minimization of proxy cost functions. The proposed solution is analysed, compared with optimal solutions and evaluated on several synthetic and real-world datasets. On these datasets, the method allows a significant improvement ($23-86\%$) in annotation efficiency.

Xavier Bou, Aitor Artola, Thibaud Ehret et al.

Detecting relevant changes is a fundamental problem of video surveillance. Because of the high variability of data and the difficulty of properly annotating changes, unsupervised methods dominate the field. Arguably one of the most critical issues to make them practical is to reduce their false alarm rate. In this work, we develop a method-agnostic weakly supervised a-contrario validation process, based on high dimensional statistical modeling of deep features, to reduce the number of false alarms of any change detection algorithm. We also raise the insufficiency of the conventionally used pixel-wise evaluation, as it fails to precisely capture the performance needs of most real applications. For this reason, we complement pixel-wise metrics with object-wise metrics and evaluate the impact of our approach at both pixel and object levels, on six methods and several sequences from different datasets. Experimental results reveal that the proposed a-contrario validation is able to largely reduce the number of false alarms at both pixel and object levels.

Xavier Bou, Elliot Vincent, Gabriele Facciolo et al.

Semantic change detection in remote sensing aims to identify land cover changes between bi-temporal image pairs. Progress in this area has been limited by the scarcity of annotated datasets, as pixel-level annotation is costly and time-consuming. To address this, recent methods leverage synthetic data or generate artificial change pairs, but out-of-domain generalization remains limited. In this work, we introduce a weak temporal supervision strategy that leverages additional temporal observations of existing single-temporal datasets, without requiring any new annotations. Specifically, we extend single-date remote sensing datasets with new observations acquired at different times and train a change detection model by assuming that real bi-temporal pairs mostly contain no change, while pairing images from different locations to generate change examples. To handle the inherent noise in these weak labels, we employ an object-aware change map generation and an iterative refinement process. We validate our approach on extended versions of the FLAIR and IAILD aerial datasets, achieving strong zero-shot and low-data regime performance across different benchmarks. Lastly, we showcase results over large areas in France, highlighting the scalability potential of our method.

Wei Tang, Jinpei Han, Kangning Cui et al.

Electrocardiogram (ECG) foundation models pretrained on typical diagnostic 10-second ECG segments, have demonstrated strong transferability across a range of clinical applications. However, many real-world applications produce recordings that are typically longer, and are varied in duration during inference time. These 10-second models have no built-in way to combine information across time. Extending them to longer horizons introduces two challenges: structural incompatibilities arising from input-length disparities, and semantic challenges that limit meaningful temporal aggregation. We propose a parameter-efficient framework that extends pretrained ECG foundation models to longer and variable-length ECGs without retraining the backbone. Guided by a frozen pretrained 10-second model, we introduce a lightweight plug-in module that extends the model in two complementary ways: (i) structurally compatible long-sequence processing and (ii) semantically informed temporal modeling. Experiments on multiple long-horizon ECG tasks, datasets, and foundation model backbones demonstrate that our method enables robust long-horizon extension from pretrained snapshot models, consistently outperforming sliding-window and pooling-based baselines with strong parameter efficiency.

Kangning Cui, Surendra Bohara, Suraj Prasai et al.

Illegal gold mining in the Amazon rainforest causes deforestation, water contamination, and long-term ecosystem disruption, yet remains difficult to monitor at fine spatial scales. Satellite imagery supports large-scale observation, but often misses small mining-related structures and subtle land-cover transitions, especially under frequent cloud cover. We introduce ELDOR, a large-scale UAV benchmark for monitoring environmental and landscape disturbance from illegal gold mining in the rainforest. ELDOR contains manually annotated orthomosaic imagery covering over 2,500 hectares, with pixel-level semantic labels for both mining-related activities and surrounding ecological structures. With this unified annotation source, we establish four benchmark tasks: semantic segmentation, segmentation-derived recognition, direct multi-label classification, and class-presence recognition with vision-language models. Across these tasks, we compare generic and remote-sensing-specific segmentation models, vision foundation model-related segmentation methods, direct multi-label classification methods, and vision-language models under a controlled closed-set protocol. Results show that current methods still struggle with rare small-scale mining structures and fine-grained recovery classes, suggesting the need for context-aware and multimodal modeling. To support domain analysis and practical use, we further build an interactive explorer for domain experts that provides a unified interface for data exploration and model inference.

Yaofang Liu, Yumeng Ren, Aitor Artola et al.

The rapid advancement of video diffusion models has been hindered by fundamental limitations in temporal modeling, particularly the rigid synchronization of frame evolution imposed by conventional scalar timestep variables. While task-specific adaptations and autoregressive models have sought to address these challenges, they remain constrained by computational inefficiency, catastrophic forgetting, or narrow applicability. In this work, we present Pusa, a groundbreaking paradigm that leverages vectorized timestep adaptation (VTA) to enable fine-grained temporal control within a unified video diffusion framework. Besides, VTA is a non-destructive adaptation, which means it fully preserves the capabilities of the base model. By finetuning the SOTA Wan2.1-T2V-14B model with VTA, we achieve unprecedented efficiency -- surpassing the performance of Wan-I2V-14B with $\leq$ 1/200 of the training cost (\$500 vs. $\geq$ \$100,000) and $\leq$ 1/2500 of the dataset size (4K vs. $\geq$ 10M samples). Pusa not only sets a new standard for image-to-video (I2V) generation, achieving a VBench-I2V total score of 87.32\% (vs. 86.86\% of Wan-I2V-14B), but also unlocks many zero-shot multi-task capabilities such as start-end frames and video extension -- all without task-specific training. Meanwhile, Pusa can still perform text-to-video generation. Mechanistic analyses reveal that our approach preserves the foundation model's generative priors while surgically injecting temporal dynamics, avoiding the combinatorial explosion inherent to vectorized timesteps. This work establishes a scalable, efficient, and versatile paradigm for next-generation video synthesis, democratizing high-fidelity video generation for research and industry alike. Code is open-sourced at https://github.com/Yaofang-Liu/Pusa-VidGen

Yaofang Liu, Kangning Cui, Meng Chu et al.

Humans often specify and create through visual artifacts: typography sheets, sketches, reference images, and annotated scenes. Yet modern visual generators still ask users to serialize this intent into text, a bottleneck that compresses signals like spatial structure, exact appearance, and glyph shape. We propose \textbf{\emph{visual-to-visual} (V2V)} generation, in which the user conditions a generative model with a visual specification page rather than a text prompt. The page is not an edit target, but a visual document that specifies the desired output. We introduce \textbf{V2V-Zero}, a training-free framework that exposes this interface in existing vision-language model (VLM) conditioned generators by replacing text-only conditioning with final-layer hidden states extracted from visual pages, exploiting the fact that the frozen VLM already maps both text and images into the generator's conditioning space. On GenEval, V2V-Zero reaches 0.85 with a frozen Qwen-Image backbone, closely matching its optimized text-to-image performance without fine-tuning. To evaluate the broader V2V space, we introduce \textbf{Simple-V2V Bench}, spanning seven visual-conditioning tasks and seven models, including GPT Image 2, Nano Banana 2, Seedream 5.0 Lite, open-weight baselines, and a video extension. V2V-Zero scores 32.7/100, outperforming evaluated open-weight image baselines and revealing a clear capability hierarchy: attribute binding is strong, content generation is unreliable, and structural control remains hard even for commercial systems. A HunyuanVideo-1.5 extension scores 20.2/100, showing the interface transfers beyond images. Mechanistic analysis shows the default reasoning path is primarily visually routed, with 95.0\% of conditioning-token attention mass on visual-page hidden states.

Xavier Bou, Gabriele Facciolo, Rafael Grompone von Gioi et al.

Oriented object detection predicts orientation in addition to object location and bounding box. Precisely predicting orientation remains challenging due to angular periodicity, which introduces boundary discontinuity issues and symmetry ambiguities. Inspired by classical works on edge and corner detection, this paper proposes to represent orientation in oriented bounding boxes as a structure tensor. This representation combines the strengths of Gaussian-based methods and angle-coder solutions, providing a simple yet efficient approach that is robust to angular periodicity issues without additional hyperparameters. Extensive evaluations across five datasets demonstrate that the proposed structure tensor representation outperforms previous methods in both fully-supervised and weakly supervised tasks, achieving high precision in angular prediction with minimal computational overhead. Thus, this work establishes structure tensors as a robust and modular alternative for encoding orientation in oriented object detection. We make our code publicly available, allowing for seamless integration into existing object detectors.

Jin Guo, Roy Y. He, Jean-Michel Morel

In 2020 Domingos introduced an interpolation formula valid for "every model trained by gradient descent". He concluded that such models behave approximately as kernel machines. In this work, we extend the Domingos formula to stochastic training. We introduce a stochastic gradient kernel that extends the deterministic version via a continuous-time diffusion approximation. We prove stochastic Domingos theorems and show that the expected network output admits a kernel-machine representation with optimizer-specific weighting. It reveals that training samples contribute through loss-dependent weights and gradient alignment along the training trajectory. We then link the generalization error to the null space of the integral operator induced by the stochastic gradient kernel. The same path-kernel viewpoint provides a unified interpretation of diffusion models and GANs: diffusion induces stage-wise, noise-localized corrections, whereas GANs induce distribution-guided corrections shaped by discriminator geometry. We visualize the evolution of implicit kernels during optimization and quantify out-of-distribution behaviors through a series of numerical experiments. Our results support a feature-space memory view of learning: training stores data-dependent information in an evolving tangent feature geometry, and predictions at test time arise from kernel-weighted retrieval and aggregation of these stored features, with generalization governed by alignment between test points and the learned feature memory.

Roy Y. He, Sung Ha Kang, Jean-Michel Morel

Image vectorization converts raster images into vector graphics composed of regions separated by curves. Typical vectorization methods first define the regions by grouping similar colored regions via color quantization, then approximate their boundaries by Bezier curves. In that way, the raster input is converted into an SVG format parameterizing the regions' colors and the Bezier control points. This compact representation has many graphical applications thanks to its universality and resolution-independence. In this paper, we remark that image vectorization is nothing but an image segmentation, and that it can be built by fine to coarse region merging. Our analysis of the problem leads us to propose a vectorization method alternating region merging and curve smoothing. We formalize the method by alternate operations on the dual and primal graph induced from any domain partition. In that way, we address a limitation of current vectorization methods, which separate the update of regional information from curve approximation. We formalize region merging methods by associating them with various gain functionals, including the classic Beaulieu-Goldberg and Mumford-Shah functionals. More generally, we introduce and compare region merging criteria involving region number, scale, area, and internal standard deviation. We also show that the curve smoothing, implicit in all vectorization methods, can be performed by the shape-preserving affine scale space. We extend this flow to a network of curves and give a sufficient condition for the topological preservation of the segmentation. The general vectorization method that follows from this analysis shows explainable behaviors, explicitly controlled by a few intuitive parameters. It is experimentally compared to state-of-the-art software and proved to have comparable or superior fidelity and cost efficiency.

Yohann Tendero, Jerome Gilles, Stephane Landeau et al.

This paper introduces a new way to correct the non-uniformity (NU) in uncooled infrared-type images. The main defect of these uncooled images is the lack of a column (resp. line) time-dependent cross-calibration, resulting in a strong column (resp. line) and time dependent noise. This problem can be considered as a 1D flicker of the columns inside each frame. Thus, classic movie deflickering algorithms can be adapted, to equalize the columns (resp. the lines). The proposed method therefore applies to the series formed by the columns of an infrared image a movie deflickering algorithm. The obtained single image method works on static images, and therefore requires no registration, no camera motion compensation, and no closed aperture sensor equalization. Thus, the method has only one camera dependent parameter, and is landscape independent. This simple method will be compared to a state of the art total variation single image correction on raw real and simulated images. The method is real time, requiring only two operations per pixel. It involves no test-pattern calibration and produces no "ghost artifacts".

Yu Guo, Axel Davy, Gabriele Facciolo et al.

With the widespread application of convolutional neural networks (CNNs), the traditional model based denoising algorithms are now outperformed. However, CNNs face two problems. First, they are computationally demanding, which makes their deployment especially difficult for mobile terminals. Second, experimental evidence shows that CNNs often over-smooth regular textures present in images, in contrast to traditional non-local models. In this letter, we propose a solution to both issues by combining a nonlocal algorithm with a lightweight residual CNN. This solution gives full latitude to the advantages of both models. We apply this framework to two GPU implementations of classic nonlocal algorithms (NLM and BM3D) and observe a substantial gain in both cases, performing better than the state-of-the-art with low computational requirements. Our solution is between 10 and 20 times faster than CNNs with equivalent performance and attains higher PSNR. In addition the final method shows a notable gain on images containing complex textures like the ones of the MIT Moire dataset.

Kangning Cui, Rongkun Zhu, Manqi Wang et al.

Palms are ecologically and economically indicators of tropical forest health, biodiversity, and human impact that support local economies and global forest product supply chains. While palm detection in plantations is well-studied, efforts to map naturally occurring palms in dense forests remain limited by overlapping crowns, uneven shading, and heterogeneous landscapes. We develop PRISM (Processing, Inference, Segmentation, and Mapping), a flexible pipeline for detecting and localizing palms in dense tropical forests using large orthomosaic images. Orthomosaics are created from thousands of aerial images and spanning several to hundreds of gigabytes. Our contributions are threefold. First, we construct a large UAV-derived orthomosaic dataset collected across 21 ecologically diverse sites in western Ecuador, annotated with 8,830 bounding boxes and 5,026 palm center points. Second, we evaluate multiple state-of-the-art object detectors based on efficiency and performance, integrating zero-shot SAM 2 as the segmentation backbone, and refining the results for precise geographic mapping. Third, we apply calibration methods to align confidence scores with IoU and explore saliency maps for feature explainability. Though optimized for palms, PRISM is adaptable for identifying other natural objects, such as eastern white pines. Future work will explore transfer learning for lower-resolution datasets (0.5 to 1m).

Xavier Bou, Gabriele Facciolo, Rafael Grompone von Gioi et al.

The goal of this paper is to perform object detection in satellite imagery with only a few examples, thus enabling users to specify any object class with minimal annotation. To this end, we explore recent methods and ideas from open-vocabulary detection for the remote sensing domain. We develop a few-shot object detector based on a traditional two-stage architecture, where the classification block is replaced by a prototype-based classifier. A large-scale pre-trained model is used to build class-reference embeddings or prototypes, which are compared to region proposal contents for label prediction. In addition, we propose to fine-tune prototypes on available training images to boost performance and learn differences between similar classes, such as aircraft types. We perform extensive evaluations on two remote sensing datasets containing challenging and rare objects. Moreover, we study the performance of both visual and image-text features, namely DINOv2 and CLIP, including two CLIP models specifically tailored for remote sensing applications. Results indicate that visual features are largely superior to vision-language models, as the latter lack the necessary domain-specific vocabulary. Lastly, the developed detector outperforms fully supervised and few-shot methods evaluated on the SIMD and DIOR datasets, despite minimal training parameters.

Bruno Galerne, Jianling Wang, Lara Raad et al.

Applying style transfer to a full 3D environment is a challenging task that has seen many developments since the advent of neural rendering. 3D Gaussian splatting (3DGS) has recently pushed further many limits of neural rendering in terms of training speed and reconstruction quality. This work introduces SGSST: Scaling Gaussian Splatting Style Transfer, an optimization-based method to apply style transfer to pretrained 3DGS scenes. We demonstrate that a new multiscale loss based on global neural statistics, that we name SOS for Simultaneously Optimized Scales, enables style transfer to ultra-high resolution 3D scenes. Not only SGSST pioneers 3D scene style transfer at such high image resolutions, it also produces superior visual quality as assessed by thorough qualitative, quantitative and perceptual comparisons.

Kangning Cui, Wei Tang, Rongkun Zhu et al.

Understanding the spatial distribution of palms within tropical forests is essential for effective ecological monitoring, conservation strategies, and the sustainable integration of natural forest products into local and global supply chains. However, the analysis of remotely sensed data in these environments faces significant challenges, such as overlapping palm and tree crowns, uneven shading across the canopy surface, and the heterogeneous nature of the forest landscapes, which often affect the performance of palm detection and segmentation algorithms. To overcome these issues, we introduce PalmDSNet, a deep learning framework for real-time detection, segmentation, and counting of canopy palms. Additionally, we employ a bimodal reproduction algorithm that simulates palm spatial propagation to further enhance the understanding of these point patterns using PalmDSNet's results. We used UAV-captured imagery to create orthomosaics from 21 sites across western Ecuadorian tropical forests, covering a gradient from the everwet Chocó forests near Colombia to the drier forests of southwestern Ecuador. Expert annotations were used to create a comprehensive dataset, including 7,356 bounding boxes on image patches and 7,603 palm centers across five orthomosaics, encompassing a total area of 449 hectares. By combining PalmDSNet with the bimodal reproduction algorithm, which optimizes parameters for both local and global spatial variability, we effectively simulate the spatial distribution of palms in diverse and dense tropical environments, validating its utility for advanced applications in tropical forest monitoring and remote sensing analysis.

Natalia Bottaioli, Solène Tarride, Jérémy Anger et al.

This study evaluates the recently proposed Document Attention Network (DAN) for extracting key-value information from Uruguayan birth certificates, handwritten in Spanish. We investigate two annotation strategies for automatically transcribing handwritten documents, fine-tuning DAN with minimal training data and annotation effort. Experiments were conducted on two datasets containing the same images (201 scans of birth certificates written by more than 15 different writers) but with different annotation methods. Our findings indicate that normalized annotation is more effective for fields that can be standardized, such as dates and places of birth, whereas diplomatic annotation performs much better for fields containing names and surnames, which can not be standardized.

Chu Chen, Aitor Artola, Yang Liu et al.

Chemical Exchange Saturation Transfer (CEST) MRI enables molecular-level visualization of low-concentration metabolites by leveraging proton exchange dynamics. However, its clinical translation is hindered by inherent challenges: spatially varying noise arising from hardware limitations, and complex imaging protocols introduce heteroscedasticity in CEST data, perturbing the accuracy of quantitative contrast mapping such as amide proton transfer (APT) imaging. Traditional denoising methods are not designed for this complex noise and often alter the underlying information that is critical for biomedical analysis. To overcome these limitations, we propose a new Blind Adaptive Local Denoising (BALD) method. BALD exploits the self-similar nature of CEST data to derive an adaptive variance-stabilizing transform that equalizes the noise distributions across CEST pixels without prior knowledge of noise characteristics. Then, BALD performs two-stage denoising on a linear transformation of data to disentangle molecular signals from noise. A local SVD decomposition is used as a linear transform to prevent spatial and spectral denoising artifacts. We conducted extensive validation experiments on multiple phantoms and \textit{in vivo} CEST scans. In these experiments, BALD consistently outperformed state-of-the-art CEST denoisers in both denoising metrics and downstream tasks such as molecular concentration maps estimation and cancer detection.

Chu Chen, Ander Biguri, Jean-Michel Morel et al.

X-ray Computed Laminography (CL) is essential for non-destructive inspection of plate-like structures in applications such as microchips and composite battery materials, where traditional computed tomography (CT) struggles due to geometric constraints. However, reconstructing high-quality volumes from laminographic projections remains challenging, particularly under highly sparse-view acquisition conditions. In this paper, we propose a reconstruction algorithm, namely LamiGauss, that combines Gaussian Splatting radiative rasterization with a dedicated detector-to-world transformation model incorporating the laminographic tilt angle. LamiGauss leverages an initialization strategy that explicitly filters out common laminographic artifacts from the preliminary reconstruction, preventing redundant Gaussians from being allocated to false structures and thereby concentrating model capacity on representing the genuine object. Our approach effectively optimizes directly from sparse projections, enabling accurate and efficient reconstruction with limited data. Extensive experiments on both synthetic and real datasets demonstrate the effectiveness and superiority of the proposed method over existing techniques. LamiGauss uses only 3$\%$ of full views to achieve superior performance over the iterative method optimized on a full dataset.

Rongkun Zhu, Kangning Cui, Wei Tang et al.

Accurate mapping of individual trees is essential for ecological monitoring and forest management. Orthomosaic imagery from unmanned aerial vehicles (UAVs) is widely used, but stitching artifacts and heavy preprocessing limit its suitability for field deployment. This study explores the use of raw UAV imagery for palm detection and crown-center localization in tropical forests. Two research questions are addressed: (1) how detection performance varies across orthomosaic and raw imagery, including within-domain and cross-domain transfer, and (2) to what extent crown-center annotations improve localization accuracy beyond bounding-box centroids. Using state-of-the-art detectors and keypoint models, we show that raw imagery yields superior performance in deployment-relevant scenarios, while orthomosaics retain value for robust cross-domain generalization. Incorporating crown-center annotations in training further improves localization and provides precise tree positions for downstream ecological analyses. These findings offer practical guidance for UAV-based biodiversity and conservation monitoring.

Diego Belzarena, Seginus Mowlavi, Aitor Artola et al.

Current OCR systems are based on deep learning models trained on large amounts of data. Although they have shown some ability to generalize to unseen data, especially in detection tasks, they can struggle with recognizing low-quality data. This is particularly evident for printed documents, where intra-domain data variability is typically low, but inter-domain data variability is high. In that context, current OCR methods do not fully exploit each document's redundancy. We propose an unsupervised method by leveraging the redundancy of character shapes within a document to correct imperfect outputs of a given OCR system and suggest better clustering. To this aim, we introduce an extended Gaussian Mixture Model (GMM) by alternating an Expectation-Maximization (EM) algorithm with an intra-cluster realignment process and normality statistical testing. We demonstrate improvements in documents with various levels of degradation, including recovered Uruguayan military archives and 17th to mid-20th century European newspapers.

Han Zhang, Xue-Cheng Tai, Jean-Michel Morel et al.

Blood flow imaging provides important information for hemodynamic behavior within the vascular system and plays an essential role in medical diagnosis and treatment planning. However, obtaining high-quality flow images remains a significant challenge. In this work, we address the problem of denoising flow images that may suffer from artifacts due to short acquisition times or device-induced errors. We formulate this task as an optimization problem, where the objective is to minimize the discrepancy between the modeled velocity field, constrained to satisfy the Navier-Stokes equations, and the observed noisy velocity data. To solve this problem, we decompose it into two subproblems: a fluid subproblem and a geometry subproblem. The fluid subproblem leverages a Physics-Informed Neural Network to reconstruct the velocity field from noisy observations, assuming a fixed domain. The geometry subproblem aims to infer the underlying flow region by optimizing a quasi-conformal mapping that deforms a reference domain. These two subproblems are solved in an alternating Gauss-Seidel fashion, iteratively refining both the velocity field and the domain. Upon convergence, the framework yields a high-quality reconstruction of the flow image. We validate the proposed method through experiments on synthetic flow data in a converging channel geometry under varying levels of Gaussian noise, and on real-like flow data in an aortic geometry with signal-dependent noise. The results demonstrate the effectiveness and robustness of the approach. Additionally, ablation studies are conducted to assess the influence of key hyperparameters.

Luis Alvarez, Jean-Michel Morel

In this paper, we propose an active contour model for silhouette vectorization using cubic Bézier curves. Among the end points of the Bézier curves, we distinguish between corner and regular points where the orientation of the tangent vector is prescribed. By minimizing the distance of the Bézier curves to the silhouette boundary, the active contour model optimizes the location of the Bézier curves end points, the orientation of the tangent vectors in the regular points, and the estimation of the Bézier curve parameters. This active contour model can use the silhouette vectorization obtained by any method as an initial guess. The proposed method significantly reduces the average distance between the silhouette boundary and its vectorization obtained by the world-class graphic software Inkscape, Adobe Illustrator, and a curvature-based vectorization method, which we introduce for comparison. Our method also allows us to impose additional regularity on the Bézier curves by reducing their lengths.

Yumeng Ren, Yaofang Liu, Aitor Artola et al.

Diffusion denoising models have become a popular approach for image generation, but they often suffer from slow convergence during training. In this paper, we identify that this slow convergence is partly due to the complexity of the Brownian motion driving the forward-time process. To address this, we represent the Brownian motion using the Karhunen-Loève expansion, truncating it to a limited number of eigenfunctions. We propose a novel ordinary differential equation with augmented random initials, termed KL diffusion, as a new forward-time process for training and sampling. By developing an appropriate denoising loss function, we facilitate the integration of our KL-diffusion into existing denoising-based models. Using the widely adopted DDIM framework as our baseline ensures a fair comparison, as our modifications focus solely on the forward process and loss function, leaving the network architecture and sampling methods unchanged. Our method significantly outperforms baseline diffusion models, achieving convergence speeds that are twice faster to reach the best FID score of the baseline and ultimately yielding much lower FID scores. Notably, our approach allows for highly parallelized computation, requires no additional learnable parameters, and can be flexibly integrated into existing diffusion methods. The code will be made publicly available.

Thomas Eboli, Jean-Michel Morel, Gabriele Facciolo

Blurry images usually exhibit similar blur at various locations across the image domain, a property barely captured in nowadays blind deblurring neural networks. We show that when extracting patches of similar underlying blur is possible, jointly processing the stack of patches yields superior accuracy than handling them separately. Our collaborative scheme is implemented in a neural architecture with a pooling layer on the stack dimension. We present three practical patch extraction strategies for image sharpening, camera shake removal and optical aberration correction, and validate the proposed approach on both synthetic and real-world benchmarks. For each blur instance, the proposed collaborative strategy yields significant quantitative and qualitative improvements.

Adrien Courtois, Damien Scieur, Jean-Michel Morel et al.

We propose SING (StabIlized and Normalized Gradient), a plug-and-play technique that improves the stability and generalization of the Adam(W) optimizer. SING is straightforward to implement and has minimal computational overhead, requiring only a layer-wise standardization of the gradients fed to Adam(W) without introducing additional hyper-parameters. We support the effectiveness and practicality of the proposed approach by showing improved results on a wide range of architectures, problems (such as image classification, depth estimation, and natural language processing), and in combination with other optimizers. We provide a theoretical analysis of the convergence of the method, and we show that by virtue of the standardization, SING can escape local minima narrower than a threshold that is inversely proportional to the network's depth.

Roland Akiki, Roger Marí, Carlo de Franchis et al.

The Rational Polynomial Camera (RPC) model can be used to describe a variety of image acquisition systems in remote sensing, notably optical and Synthetic Aperture Radar (SAR) sensors. RPC functions relate 3D to 2D coordinates and vice versa, regardless of physical sensor specificities, which has made them an essential tool to harness satellite images in a generic way. This article describes a terrain-independent algorithm to accurately derive a RPC model from a set of 3D-2D point correspondences based on a regularized least squares fit. The performance of the method is assessed by varying the point correspondences and the size of the area that they cover. We test the algorithm on SAR and optical data, to derive RPCs from physical sensor models or from other RPC models after composition with corrective functions.

Yu Guo, Qiyu Jin, Gabriele Facciolo et al.

Image demosaicking and denoising are the first two key steps of the color image production pipeline. The classical processing sequence has for a long time consisted of applying denoising first, and then demosaicking. Applying the operations in this order leads to oversmoothing and checkerboard effects. Yet, it was difficult to change this order, because once the image is demosaicked, the statistical properties of the noise are dramatically changed and hard to handle by traditional denoising models. In this paper, we address this problem by a hybrid machine learning method. We invert the traditional color filter array (CFA) processing pipeline by first demosaicking and then denoising. Our demosaicking algorithm, trained on noiseless images, combines a traditional method and a residual convolutional neural network (CNN). This first stage retains all known information, which is the key point to obtain faithful final results. The noisy demosaicked image is then passed through a second CNN restoring a noiseless full-color image. This pipeline order completely avoids checkerboard effects and restores fine image detail. Although CNNs can be trained to solve jointly demosaicking-denoising end-to-end, we find that this two-stage training performs better and is less prone to failure. It is shown experimentally to improve on the state of the art, both quantitatively and in terms of visual quality.

Qiyu Jin, Gabriele Facciolo, Jean-Michel Morel

Image denoising and demosaicking are the most important early stages in digital camera pipelines. They constitute a severely ill-posed problem that aims at reconstructing a full color image from a noisy color filter array (CFA) image. In most of the literature, denoising and demosaicking are treated as two independent problems, without considering their interaction, or asking which should be applied first. Several recent works have started addressing them jointly in works that involve heavy weight CNNs, thus incompatible with low power portable imaging devices. Hence, the question of how to combine denoising and demosaicking to reconstruct full color images remains very relevant: Is denoising to be applied first, or should that be demosaicking first? In this paper, we review the main variants of these strategies and carry-out an extensive evaluation to find the best way to reconstruct full color images from a noisy mosaic. We conclude that demosaicking should applied first, followed by denoising. Yet we prove that this requires an adaptation of classic denoising algorithms to demosaicked noise, which we justify and specify.

Axel Davy, Thibaud Ehret, Jean-Michel Morel et al.

Anomaly detectors address the difficult problem of detecting automatically exceptions in an arbitrary background image. Detection methods have been proposed by the thousands because each problem requires a different background model. By analyzing the existing approaches, we show that the problem can be reduced to detecting anomalies in residual images (extracted from the target image) in which noise and anomalies prevail. Hence, the general and impossible background modeling problem is replaced by simpler noise modeling, and allows the calculation of rigorous thresholds based on the a contrario detection theory. Our approach is therefore unsupervised and works on arbitrary images.

Thibaud Ehret, Axel Davy, Jean-Michel Morel et al.

Modeling the processing chain that has produced a video is a difficult reverse engineering task, even when the camera is available. This makes model based video processing a still more complex task. In this paper we propose a fully blind video denoising method, with two versions off-line and on-line. This is achieved by fine-tuning a pre-trained AWGN denoising network to the video with a novel frame-to-frame training strategy. Our denoiser can be used without knowledge of the origin of the video or burst and the post processing steps applied from the camera sensor. The on-line process only requires a couple of frames before achieving visually-pleasing results for a wide range of perturbations. It nonetheless reaches state of the art performance for standard Gaussian noise, and can be used off-line with still better performance.

Axel Davy, Thibaud Ehret, Jean-Michel Morel et al.

Non-local patch based methods were until recently state-of-the-art for image denoising but are now outperformed by CNNs. Yet they are still the state-of-the-art for video denoising, as video redundancy is a key factor to attain high denoising performance. The problem is that CNN architectures are hardly compatible with the search for self-similarities. In this work we propose a new and efficient way to feed video self-similarities to a CNN. The non-locality is incorporated into the network via a first non-trainable layer which finds for each patch in the input image its most similar patches in a search region. The central values of these patches are then gathered in a feature vector which is assigned to each image pixel. This information is presented to a CNN which is trained to predict the clean image. We apply the proposed architecture to image and video denoising. For the latter patches are searched for in a 3D spatio-temporal volume. The proposed architecture achieves state-of-the-art results. To the best of our knowledge, this is the first successful application of a CNN to video denoising.

Charles Hessel, Jean-Michel Morel

This paper introduces a quantitative evaluation of filters that seek to separate an image into its large-scale variations, the base layer, and its fine-scale variations, the detail layer. Such methods have proliferated with the development of HDR imaging and the proposition of many new tone-mapping operators. We argue that an objective quality measurement for all methods can be based on their artifacts. To this aim, the four main recurrent artifacts are described and mathematically characterized. Among them two are classic, the luminance halo and the staircase effect, but we show the relevance of two more, the contrast halo and the compartmentalization effect. For each of these artifacts we design a test-pattern and its attached measurement formula. Then we fuse these measurements into a single quality mark, and obtain in that way a ranking method valid for all filters performing a base+detail decomposition. This synthetic ranking is applied to seven filters representative of the literature and shown to agree with expert artifact rejection criteria.

Thibaud Ehret, Axel Davy, Jean-Michel Morel et al.

We review the broad variety of methods that have been proposed for anomaly detection in images. Most methods found in the literature have in mind a particular application. Yet we show that the methods can be classified mainly by the structural assumption they make on the "normal" image. Five different structural assumptions emerge. Our analysis leads us to reformulate the best representative algorithms by attaching to them an a contrario detection that controls the number of false positives and thus derive universal detection thresholds. By combining the most general structural assumptions expressing the background's normality with the best proposed statistical detection tools, we end up proposing generic algorithms that seem to generalize or reconcile most methods. We compare the six best representatives of our proposed classes of algorithms on anomalous images taken from classic papers on the subject, and on a synthetic database. Our conclusion is that it is possible to perform automatic anomaly detection on a single image.

José Lezama, Samy Blusseau, Jean-Michel Morel et al.

Many psychophysical studies are dedicated to the evaluation of the human gestalt detection on dot or Gabor patterns, and to model its dependence on the pattern and background parameters. Nevertheless, even for these constrained percepts, psychophysics have not yet reached the challenging prediction stage, where human detection would be quantitatively predicted by a (generic) model. On the other hand, Computer Vision has attempted at defining automatic detection thresholds. This chapter sketches a procedure to confront these two methodologies inspired in gestaltism. Using a computational quantitative version of the non-accidentalness principle, we raise the possibility that the psychophysical and the (older) gestaltist setups, both applicable on dot or Gabor patterns, find a useful complement in a Turing test. In our perceptual Turing test, human performance is compared by the scientist to the detection result given by a computer. This confrontation permits to revive the abandoned method of gestaltic games. We sketch the elaboration of such a game, where the subjects of the experiment are confronted to an alignment detection algorithm, and are invited to draw examples that will fool it. We show that in that way a more precise definition of the alignment gestalt and of its computational formulation seems to emerge. Detection algorithms might also be relevant to more classic psychophysical setups, where they can again play the role of a Turing test. To a visual experiment where subjects were invited to detect alignments in Gabor patterns, we associated a single function measuring the alignment detectability in the form of a number of false alarms (NFA). The first results indicate that the values of the NFA, as a function of all simulation parameters, are highly correlated to the human detection. This fact, that we intend to support by further experiments , might end up confirming that human alignment detection is the result of a single mechanism.

Lara Raad, Axel Davy, Agnès Desolneux et al.

Exemplar-based texture synthesis is the process of generating, from an input sample, new texture images of arbitrary size and which are perceptually equivalent to the sample. The two main approaches are statistics-based methods and patch re-arrangement methods. In the first class, a texture is characterized by a statistical signature; then, a random sampling conditioned to this signature produces genuinely different texture images. The second class boils down to a clever "copy-paste" procedure, which stitches together large regions of the sample. Hybrid methods try to combine ideas from both approaches to avoid their hurdles. The recent approaches using convolutional neural networks fit to this classification, some being statistical and others performing patch re-arrangement in the feature space. They produce impressive synthesis on various kinds of textures. Nevertheless, we found that most real textures are organized at multiple scales, with global structures revealed at coarse scales and highly varying details at finer ones. Thus, when confronted with large natural images of textures the results of state-of-the-art methods degrade rapidly, and the problem of modeling them remains wide open.

Martin Rais, Gabriele Facciolo, Enric Meinhardt-Llopis et al.

We reconsider the classic problem of estimating accurately a 2D transformation from point matches between images containing outliers. RANSAC discriminates outliers by randomly generating minimalistic sampled hypotheses and verifying their consensus over the input data. Its response is based on the single hypothesis that obtained the largest inlier support. In this article we show that the resulting accuracy can be improved by aggregating all generated hypotheses. This yields RANSAAC, a framework that improves systematically over RANSAC and its state-of-the-art variants by statistically aggregating hypotheses. To this end, we introduce a simple strategy that allows to rapidly average 2D transformations, leading to an almost negligible extra computational cost. We give practical applications on projective transforms and homography+distortion models and demonstrate a significant performance gain in both cases.

Boshra Rajaei, Rafael Grompone von Gioi, Jean-Michel Morel

In this paper, we reconsider the early computer vision bottom-up program, according to which higher level features (geometric structures) in an image could be built up recursively from elementary features by simple grouping principles coming from Gestalt theory. Taking advantage of the (recent) advances in reliable line segment detectors, we propose three feature detectors that constitute one step up in this bottom up pyramid. For any digital image, our unsupervised algorithm computes three classic Gestalts from the set of predetected line segments: good continuations, nonlocal alignments, and bars. The methodology is based on a common stochastic {\it a contrario model} yielding three simple detection formulas, characterized by their number of false alarms. This detection algorithm is illustrated on several digital images.

Ives Rey-Otero, Jean-Michel Morel, Mauricio Delbracio

The most popular image matching algorithm SIFT, introduced by D. Lowe a decade ago, has proven to be sufficiently scale invariant to be used in numerous applications. In practice, however, scale invariance may be weakened by various sources of error inherent to the SIFT implementation affecting the stability and accuracy of keypoint detection. The density of the sampling of the Gaussian scale-space and the level of blur in the input image are two of these sources. This article presents a numerical analysis of their impact on the extracted keypoints stability. Such an analysis has both methodological and practical implications, on how to compare feature detectors and on how to improve SIFT. We show that even with a significantly oversampled scale-space numerical errors prevent from achieving perfect stability. Usual strategies to filter out unstable detections are shown to be inefficient. We also prove that the effect of the error in the assumption on the initial blur is asymmetric and that the method is strongly degraded in presence of aliasing or without a correct assumption on the camera blur.

Ives Rey-Otero, Mauricio Delbracio, Jean-Michel Morel

Most computer vision application rely on algorithms finding local correspondences between different images. These algorithms detect and compare stable local invariant descriptors centered at scale-invariant keypoints. Because of the importance of the problem, new keypoint detectors and descriptors are constantly being proposed, each one claiming to perform better (or to be complementary) to the preceding ones. This raises the question of a fair comparison between very diverse methods. This evaluation has been mainly based on a repeatability criterion of the keypoints under a series of image perturbations (blur, illumination, noise, rotations, homotheties, homographies, etc). In this paper, we argue that the classic repeatability criterion is biased towards algorithms producing redundant overlapped detections. To compensate this bias, we propose a variant of the repeatability rate taking into account the descriptors overlap. We apply this variant to revisit the popular benchmark by Mikolajczyk et al., on classic and new feature detectors. Experimental evidence shows that the hierarchy of these feature detectors is severely disrupted by the amended comparator.