Thematic Workshops of Conference on Graphics, Patterns and Images

SIBGRAPI 2025

“Brazilian-French Workshop on Graph-based Image Analysis”

The Brazilian-French Workshop on Graph-based Image Analysis is composed of talks presented by researchers involved in an international cooperation between France and Brazil, initiated in 1989 by Arnaldo de Albuquerque (UFMG) and Gilles Bertrand (ESIEE Paris). In more than 30 years, the cooperation has been supported by several  agreements (CAPES and COFECUB, BRAFITEC, STICAMSUD, Invited Researcher; FAPEMIG and INRIA) involving UFMG, PUC Minas, UNICAMP, UFOP, ESIEE Paris, ENSEA Paris, IRISA Rennes, UGE, UPMC, U Caen, IT Grenoble. The talks to be presented in the Workshop represent well the research done by the partners in the last five years. Our team organized the France-Brazil Digital Video Journey which happened at SIBGAPI 2009, PUC Rio, by occasion of the 2009 French Year in Brazil.

The Workshop is proposed in the context of the Brazilian-French 200 years of  Bilateral Relations comemorations, 2025 French Year in Brazil. The chosen date for the Workshop is October 1st 2025. There will be no printed material for the presented talks.

Key words: Image segmentation, Mathematical morphology, Digital topology, Morse sequences, Graph-based algorithms, Power watershed, Hierarchical watershed.

Organizers: Jean Cousty (ESIEE Paris), Silvio Jamil Guimarães (PUC Minas), Alexandre Falcão (UNICAMP), Arnaldo de Albuquerque (UFMG, ESIEE Paris).

Contact: Alexandre Falcão (afalcao@unicamp.br), Silvio Jamil Guimarães (sjamil@pucminas.br) and Arnaldo de Albuquerque (arnaldo@dcc.ufmg.br, arnaldo.albuquerque@esiee.fr)

Invited speakers: Gilles Bertrand (ESIEE Paris), Laurent Najman (ESIEE Paris), Benjamin Perret (ESIEE Paris), Yukiko Kenmochi (U Caen), Silvio Guimarães (PUC-Minas), and Alexandre Falcão (UNICAMP).

Speaker, Title of the talk, and short biography

Speaker: Gilles Bertrand

Title: Morse sequences, a simple approach to discrete Morse theory.

Abstract:

We introduce the notion of a Morse sequence, offering an alternative approach to discrete Morse theory that is both simple and effective. A Morse sequence on a finite simplicial complex consists solely of two elementary operations: expansions (the inverse of collapses) and fillings (the inverse of perforations). Alternatively, a Morse sequence can be constructed using only collapses and perforations, providing a dual perspective. Such sequences serve as another representation of the gradient vector field of an arbitrary discrete Morse function.

To each Morse sequence, we associate a reference map and an extension map. The reference map assigns a set of critical simplices to each simplex in the complex, while the extension map assigns a set of simplices to each critical simplex. By considering the boundary of each critical simplex, these maps yield a chain complex that corresponds exactly to the Morse complex, that is a complex which captures the homology of the original object.

Short Bio:

Gilles Bertrand received his Ingénieur’s degree from the Ecole Centrale des Arts et Manufactures in 1976. Until 1983 he was with the Thomson-CSF company where he designed image processing systems for aeronautical applications. He received his Ph.D from the Ecole Centrale in 1986. He is currently doing research with the Computer Science and Telecommunication Department of ESIEE-Paris and with the Laboratoire d’Informatique Gaspard-Monge of Gustave Eiffel University. His research interests include digital and combinatorial topology, mathematical morphology, image analysis.

Speaker: Alexandre Falcão

Title: Learning node features for graph-based image analysis: From encoders to arc weights

Abstract:

Graph-based algorithms have proven to be powerful tools for imagese gmentation and classification tasks. In image segmentation, graph nodes typically represent pixels or superpixels, while in image classification, entire images serve as nodes. A critical component in both scenarios is the estimation of arc weights, which quantify the distances or similarities between nodes and directly impact algorithm performance.

Deep neural networks present an opportunity to learn more effective node representations. However, standard encoders often produce embeddings optimized for their original training objectives, which may not be suitable for estimating arc weights in graph-based algorithms.

This talk presents a study investigating how to learn node features tailored explicitly for suitable arc-weight estimation in graph-based image analysis. We explore multiple encoding strategies followed by dimensionality reduction, including pre-trained encoders, convolutional encoders trained from scratch, and encoders developed using the Feature Learning from Image Markers (FLIM) methodology, an innovative approach that creates effective representations from

minimal training data.

Our experimental framework addresses both segmentation and classification problems through Optimum-Path Forest (OPF) algorithms, including dynamic trees, discriminative, and transductive OPF classifiers. By comparing these different encoding and dimensionality reduction approaches, we analyze how design choices affect the quality of learned arc weights and, consequently, the performance of graph-based algorithms.

The discussion of these findings will foster new collaborative research directions among participants of the first French-Brazilian Workshop at SIBGRAPI.

Short bio:

Alexandre Xavier Falcão is a Professor in Computer Science at the Institute of Computing, State University of Campinas (UNICAMP). He holds a PhD from UNICAMP (1997), with a focus on medical image analysis, and was affiliated with the University of Pennsylvania from 1994 to 1996. He has been in the image analysis field for 35 years, with projects in video quality assessment (Globo TV, 1997), plant phenotyping (Cornell University, 2011-2012), and several other image analysis applications developed at UNICAMP since 1998. He holds a CNPq productivity research fellowship at level 1A and has been awarded two grants from Labex-Bézout for joint research at ESIEE/Paris in 2023 and 2024. He has authored over 360 papers and licensed over ten technologies, with five currently in the market. His research interests cover graph algorithms, image analysis, data visualization, and design of machine (deep) learning models.

Speaker: Laurent Najman

Title: Power-Watershed: a graph-based optimization framework for image and data processing

Abstract:

The watershed is a tool originally developed in the mathematical morphology community for image

segmentation. Over time, numerous adaptations have emerged, proving its applicability beyond image processing. Particularly, it has demonstrated its value in tasks such as data classification and filtering. This presentation aims to highlight the continued relevance of the watershed method in the era of deep learning.

Following an overview of cutting-edge achievements attainable through the watershed technique, our focus shifts to the Power Watershed (PW) optimization framework, within the context of image and data processing. Existing literature recognizes the PW framework's potential, especially when applied to prominent graph-based data processing algorithms like random walker, isoperimetric partitioning, ratio-cut clustering, multi-cut, and shortest path filters. The result is faster, yet consistent, solutions.

During the talk, we delve into the PW framework's underlying concept: the utilization of contrast invariance within data. This concept is elucidated using hypothetical images and experimental demonstrations. Importantly, the principles from this framework can be adapted for a wide range of graph-based cost minimization methods, and seamlessly integrated with deep learning networks, leading to scalable algorithms that consistently deliver high-quality outcomes.

Short bio:

Laurent Najman received the Habilitation à Diriger les Recherches in 2006 from the University of Marne-la-Vallée, a Ph.D. in applied mathematics from Paris-Dauphine University in 1994 with the highest honor (Félicitations du Jury) and an “Ingénieur” degree from the Ecole des Mines de Paris in 1991. After earning his engineering degree, he worked in the Central Research Laboratories of Thomson-CSF for three years, working on some problems of infrared image segmentation using mathematical morphology. He then joined a start-up company named Animation Science in 1995, as director of research and development. The technology of particle systems for computer graphics and scientific visualization, developed by the company under his technical leadership received several awards, including the “European Information Technology Prize 1997” awarded by the European Commission (Esprit program) and by the European Council for Applied Science and Engineering and the “Hottest Products of the Year 1996” awarded by the Computer Graphics World journal. In 1998, he joined OCÉ Print Logic Technologies, as senior scientist. He worked there on various problem of image analysis dedicated to scanning and printing. In 2002, he joined the Computer Sciences Department of ESIEE, Paris, where he is full professor and the leader of the A3SI team of the Laboratoire d’Informatique Gaspard Monge, Université Gustave Eiffel. He is currently a professor at Khalifa University, on leave from Université Gustave Eiffel. His current research interests include the study of the topology of discrete structures (such as graphs, hierarchies, and simplicial complexes), using discrete mathematical morphology and discrete optimization.

Speaker: Benjamin Perret

Title: Playing with Kruskal: algorithms for flat and hierarchical watershed cuts

Abstract:

Within the context of edge-weighted graphs, watersheds — a traditional tool for image segmentation — have been shown to be connected to well-known optimization problems, such as the minimum spanning tree problem. This has enabled the development of efficient algorithms for computing hierarchical watershed segmentations. In this talk, we will present a unified, end-to-end pipeline of algorithms for computing (hierarchical) watershed segmentations. This pipeline starts with the computation of graph-based image representations and ends with the computation of the connected components of the final (hierarchical) segmentation. This pipeline will be demonstrated using the Higra Python library for hierarchical graph analysis.

Short bio:

Benjamin Perret received the Ph.D. degree in image processing from Université de Strasbourg, France, in 2007 and the Habilitation degree from Université Paris-Est in 2021. He currently holds a Professor position at ESIEE Paris, Université Gustave Eiffel, affiliated with the Laboratoire d’Informatique Gaspard Monge. His current research interests include image analysis, machine learning and mathematical morphology.

Speakers: Yukiko Kenmochi and Silvio Guimarães

Title: Topological image analysis based on morphological hierarchies

Abstract:

Understanding and manipulating the structural and topological properties of images is essential for robust and interpretable analysis in many image processing and computer vision tasks. This talk presents a unified framework for topological image analysis based on morphological hierarchies, with a focus on component trees, the tree of shapes, and the topological tree of shapes. These hierarchical representations provide compact, scale-aware, and topology-preserving descriptions of image content, enabling structure-aware filtering, segmentation, and feature extraction.

We begin with a conceptual overview of how digital topology and mathematical morphology can be used to characterize features such as connected components, holes, and tunnels. We then introduce component trees (such as max-trees and min-trees) and the tree of shapes, a topologically complete and contrast-invariant representation that captures object-level shapes in a hierarchical form as well as the topological tree of shapes, a topologically compressed hierarchical representation of tree of shapes.

In parallel, persistent homology, a key tool in topological data analysis (TDA), has emerged as a powerful method for studying the shape of data across multiple scales. Within image analysis, persistent homology helps track the evolution of topological features as intensity thresholds vary. We examine how persistent homology relates to morphological hierarchies and how their interplay deepens our understanding of image topology. This connection has also inspired the development of new topological loss functions based on morphological structures. Some examples of medical image applications will be presented.
 
In the latter part of the talk, we discuss recent advances in topological loss functions, including those derived from persistent homology and component tree-based formulations. These loss functions have been successfully applied in deep learning to enforce topological correctness—such as maintaining connectivity or preserving the number of holes—in segmentation and reconstruction tasks.

This talk is designed for students, researchers, and practitioners who are interested in combining topological concepts with practical image analysis methods, and for those aiming to bridge the gap between geometry, topology, and machine learning.

Short bio:

Yukiko Kenmochi holds a D.Eng degree in Computer Science from Chiba University, Japan (1998) and a Habilitation degree from Université Paris-Est, France (2018). Her career started in Japan as a Research Associate at JAIST (1998-2003) and a Lecturer at Okayama University (2003-2004). During this time, she also held a JSPS Overseas Research Fellowship for staying at ESIEE Paris, France (2000-2002). In 2004, she joined the CNRS as a researcher at the Laboratoire d'informatique Gaspard-Monge, France, and moved to the Laboratoire GREYC in 2021. Her primary research interests include digital geometry and topology for computer imagery. She has also recently focused on mathematical morphology.

Silvio Jamil Ferzoli Guimarães holds a Ph.D. in Computer Science from the Federal University of Minas Gerais (2003) and a joint Ph.D. in Informatique - Université de Marne La Vallée (2003). He is currently Associate Professor IV at the Pontifical Catholic University of Minas Gerais (PUC Minas) and an Associate Researcher at ESIEE / Paris. He holds a research productivity fellowship at level 2 from CNPq. He has experience in Computer Science, working mainly in digital video analysis and processing, mathematical morphology, graph-based image and video processing, digital image processing, and information retrieval. Regarding superpixels, Prof. Silvio is co-author of several papers involving this subject and has supervised several undergraduate and graduate theses.

Workshop Organizers - Short Biographies

Jean Cousty received the engineering degree from ESIEE Paris, France in 2004, the Ph.D. degree from Université de Marne-la-Vallée in 2007 and the Habilitation à Diriger des Recherches from Université Paris-Est in 2018. After a one-year postdoctoral position with the ASCLEPIOS research team at INRIA Sophia-Antipolis, he joined the Computer Science Department at ESIEE Paris and the Laboratoire d’Informatique Gaspard-Monge at Université Gustave Eiffel, where he has been teaching and conducting research. From 2015 to 2017, he was an invited Professor in Brazil at UFMG and PUC Minas. His current research interests include graph-based approaches to image analysis and computer vision, hierarchical analysis, mathematical morphology, and discrete topology.

Silvio Jamil Ferzoli Guimarães holds a Ph.D. in Computer Science from the Federal University of Minas Gerais (2003) and a joint Ph.D. in Informatique - Université de Marne La Vallée (2003). He is currently Associate Professor IV at the Pontifical Catholic University of Minas Gerais (PUC Minas) and an Associate Researcher at ESIEE / Paris. He holds a research productivity fellowship at level 2 from CNPq. He has experience in Computer Science, working mainly in digital video analysis and processing, mathematical morphology, graph-based image and video processing, digital image processing, and information retrieval. Regarding superpixels, Prof. Silvio is co-author of several papers involving this subject and has supervised several undergraduate and graduate theses.

Alexandre Xavier Falcão is a Professor in Computer Science at the Institute of Computing, State University of Campinas (UNICAMP). He holds a PhD from UNICAMP (1997), with a focus on medical image analysis, and was affiliated with the University of Pennsylvania from 1994 to 1996. He has been in the image analysis field for 35 years, with projects in video quality assessment (Globo TV, 1997), plant phenotyping (Cornell University, 2011-2012), and several other image analysis applications developed at UNICAMP since 1998. He holds a CNPq productivity research fellowship at level 1A and has been awarded two grants from Labex-Bézout for joint research at ESIEE/Paris in 2023 and 2024. He has authored over 360 papers and licensed over ten technologies, with five currently in the market. His research interests cover graph algorithms, image analysis, data visualization, and design of machine (deep) learning models.

Arnaldo de Albuquerque Araujo  holds a D.Sc. degree in Electrical Engineering from the Universidade Federal da Paraiba (1987), Brazil, focusing on digital image processing at the Reinisch-Westifaelische Technische Hochschule Aachen, Germany, 1981-1985. In France, he did sabbatical years  in Computer Science, at ESIEE Paris, 1994-1995, ENSEA Cergy-Pontoise, 2005, and Universisité Pierre et Marie Curie - Paris 6, 2008-2009. He was an associate professor at the Electrical Engineering Department, UFPB, Brazil, 1978-1989. He is a full professor at the Computer Science Department, Universidade Federal de Minas Gerais (UFMG) (since 1990), Brazil, and invited researcher at ESIEE Paris, Université Gustave Eiffel (since 2018).  His research interests include image processing, computer vision, machine learning, multimedia information systems, and content based information (image and video) retrieval.