Graphs provide a powerful tool for coping with the non-uniformity and irregularity of 3D meshes, enabling multi-scale representations of 3D data. However, many existing methods either neglect the importance of relationships among elements within each region or fail to incorporate the hierarchical structure inherent in graph elements. This work introduces a novel representation that utilizes hierarchical segmentation techniques to address these limitations. Based on this hierarchical representation, two graph neural networks (GNNs) for 3D mesh classification are presented: (i) BM, a novel architecture that utilizes only the base graph; and (ii) HRGNet, a hierarchical model to explore multiple hierarchical levels. Experimental results demonstrate that the proposed models achieve high accuracy for the Manifold40 dataset, utilizing fewer parameters than other state-of-the-art methods. BM and HRGNet achieve results close to state-of-the-art, while BM has only 3

Salient Object Detection (SOD) methods can locate objects that stand out in an image, assign higher values to their pixels in a saliency map, and binarize the map outputting a predicted segmentation mask. A recent tendency is to investigate pre-trained lightweight models rather than deep neural networks in SOD tasks, coping with applications under limited computational resources. In this context, we have investigated lightweight networks using a methodology named Feature Learning from Image Markers (FLIM), which assumes that the encoder's kernels can be estimated from marker pixels on discriminative regions of a few representative images. This work proposes flyweight networks, hundreds of times lighter than lightweight models, for SOD by combining a FLIM encoder with an adaptive decoder, whose weights are estimated for each input image by a given heuristic function. Such FLIM networks are trained from three to four representative images only and without backpropagation, making the models suitable for applications under labeled data constraints as well. We study five adaptive decoders; two of them are introduced here. Differently from the previous ones that rely on one neuron per pixel with shared weights, the heuristic functions of the new adaptive decoders estimate the weights of each neuron per pixel. We compare FLIM models with adaptive decoders for two challenging SOD tasks with three lightweight networks from the state-of-the-art, two FLIM networks with decoders trained by backpropagation, and one FLIM network whose labeled markers define the decoder's weights. The experiments demonstrate the advantages of the proposed networks over the baselines, revealing the importance of further investigating such methods in new applications.

Capturing relevant content from videos while preserving temporal coherence remains a central challenge in video skimming, particularly in the presence of redundant and repetitive visual patterns. Conventional approaches often struggle to model temporal dependencies across different granularities, leading to suboptimal summaries that fail to balance local consistency and global structure. In this work, we propose the MUlti-ScaLe Memory-Augmented LSTM for Dynamic Video Summarization (MUSL-MALSumm), a supervised architecture based on Extended Long Short-Term Memory (xLSTM), which introduces a multi-scale temporal modeling strategy to better capture both short- and long-range dependencies. Specifically, the proposed method integrates parallel temporal convolutions with different receptive fields to encode multi-scale patterns, combined with a residual projection mechanism that preserves locally informative features while enabling effective global temporal reasoning through recurrent dynamics. This design enhances the model’s ability to maintain temporal coherence and robustness to noisy or redundant frames, while preserving a low computational footprint. Experimental results demonstrate the effectiveness of the proposed approach, achieving an average F-score of 53.92 on the SumMe dataset and 62.38 on TVSum, outperforming recent supervised baselines. Additionally, the model attains a Kendall’s τ of 0.184 and Spearman’s ρ of 0.246, indicating strong alignment with human annotations. These findings highlight the benefits of multi-scale temporal modeling in video summarization, providing a competitive and efficient solution for generating coherent and informative video summaries.

Image segmentation is a crucial step in various visual applications, including environmental monitoring through remote sensing. In the context of the ForestEyes project, which combines citizen science and machine learning to detect deforestation in tropical forests, image segments are used for labeling by volunteers and subsequent model training. Traditionally, the Simple Linear Iterative Clustering (SLIC) algorithm is adopted as the segmentation method. However, recent studies have indicated that other superpixel-based methods outperform SLIC in remote sensing image segmentation, and might suggest that they are more suitable for the task of detecting deforested areas. In this sense, this study investigated the impact of the four best segmentation methods, together with SLIC, on the training of classifiers for the target application. Initially, the results showed little variation in performance among segmentation methods, even when selecting the top five classifiers using the PyCaret AutoML library. However, by applying a classifier fusion approach (ensemble of classifiers), noticeable improvements in balanced accuracy were observed, highlighting the importance of both the choice of segmentation method and the combination of machine learning-based models for deforestation detection tasks.

Many image segmentation techniques generate sequences of region partitions by tuning parameters that control the level of detail or abstraction, but sequences often violate hierarchical consistency, where each region at one level should be entirely contained within a region at the next. Such inconsistencies undermine the benefits of hierarchical models in multiscale analysis, such as structural coherence and computational efficiency. Despite the importance of hierarchy, existing methods often overlook how to detect and correct violations in segmentation sequences. In this work, we propose a framework for identifying and classifying partitions that most harm the hierarchy. We introduce a recursive measure to quantify the accumulated inconsistency within a sequence and present an optimization strategy that pinpoints and removes the most harmful partition. Experiments on several datasets demonstrate that our approach effectively improves the hierarchiness of the sequence by analyzing and refining the segmentation series.

In the image processing field, Graph Neural Networks (GNNs) are employed to learn from the graph structure to its full potential. Before model training, the data transformation converts the image into a graph structure. In this sense, many superpixel algorithms such as SLIC, DISF, SCALP, and ODISF have been adopted for reducing complexity and creating meaningful pixel regions based on some criteria. As these algorithms offer distinct characteristics and superpixel aspects, this work-inprogress presents a study on the performance of the HiErarchical Layered Multigraph Network (HELMNet) according to the variation of superpixel methods used to generate the base graph of the HiErarchical Layered Multigraph (HELM) representation with images from the STL-10 dataset. An attention function called Region Graph Readout (RGR), used for classification, is also set to guide the clarification of the divergent results. Furthermore, this study aims to highlight the nonexistence of a general rule for selecting a superpixel method given a specific visual task.

We present SkimCap, a transformer-based video captioning framework that integrates a memory-augmented architecture with adaptive attention and a novel feature selection strategy grounded in hierarchical video skimming. Unlike traditional approaches that rely on uniformly sampled frames or pre-defined temporal segments, SkimCap performs unsupervised hierarchical clustering to identify and extract semantically salient video shots. These condensed representations provide a compact yet information-rich input to the captioning model, enabling more accurate and contextually grounded sentence generation. The memory module enhances long-range dependency modeling, while adaptive attention improves temporal alignment between visual cues and generated tokens. We evaluate SkimCap on ActivityNet, achieving CIDEr-D of 25.44, a BLEU-4 (B@4) of 10.77, and a lower Repetition-4 (R@4) score of 5.84, representing consistent caption quality and relevance improvements. An ablation study confirms the effectiveness of hierarchical skimming as a feature selection mechanism, highlighting its contribution to overall performance. SkimCap sets a new direction for incorporating structured visual summarization into end-to-end captioning systems.

Video skimming aims to generate concise yet informative summaries that highlight the most salient aspects of a video. However, conventional methods often struggle with diverse and redundant content due to their limited ability to detect scene transitions and insufficient temporal modeling. To address these challenges, we propose Streamlined Extended Long Short-Term Memory (StreamExLSTM), a supervised architecture derived from a streamlined variant of the extended Long Short-Term Memory (xLSTM) model. The proposed approach introduces two lightweight modules: ssLSTM, which captures short-range temporal dependencies through convolutional and recurrent operations, and smLSTM, which models long-range narrative structure using stacked memory-enhanced LSTMs. This dual-path design enables the model to balance local detail with global coherence while maintaining low complexity. Experimental results demonstrate that StreamExLSTM outperforms recent supervised baselines, achieving an average F-score of 48.8 on SumMe and 61.1 on TVSum. Moreover, when trained on a combined dataset, it reaches an F-score of 83.7 on the TVSum test set, performing comparably to semi-supervised, reinforcement learning, and GAN-based methods. These results validate StreamExLSTM as an effective and lightweight solution for dynamic video summarization.

Remote sensing scene classification (RSSC) poses significant challenges due to high spatial variability, complex textures, and semantic ambiguity in remote sensing imagery. While convolutional neural networks (CNNs) and transformer-based models have achieved notable success in this domain, their performance often depends on large-scale pretraining and substantial computational resources. Graph neural networks (GNNs) have emerged as a promising alternative to traditional deep learning methods by explicitly modeling the relational structure of image regions through graph representations, which have already demonstrated promising results across various image-based tasks involving images. In this work, we explore two GNN architectures tailored for RSSC: BRMv2, a novel simplified graph model built on a base region adjacency graph (RAG), and modified hierarchical layered multigraph network (mHELMNet), a modified hierarchical multigraph model that encodes multiscale and spatial relationships through a multigraph representation. Both models were evaluated on the EUROSAT and RESISC45 datasets, achieving accuracy comparable to, or in some cases exceeding, that of state-of-the-art CNN-based, hybrid GNN-based, and transformer-based methods, while using significantly fewer parameters and without relying on pretraining. Experimental results demonstrated that the proposed GNN models, mHELMNet and BRMv2, achieved over 96% accuracy on EUROSAT and approximately 85% on RESISC45, while requiring only 0.14% and 0.03% of the parameters of the leading transformer-based approach, respectively.

Visual Dialog is a challenging multimodal task requiring models to answer questions about images through multi-turn conversations. Despite significant progress, research has predominantly focused on English, limiting applicability to the 850+ million speakers of Portuguese and Spanish worldwide. We present the first comprehensive study of monolingual and multilingual Visual Dialog models for Portuguese and Spanish, introducing novel insights into cross-lingual visual grounding mechanisms. Through extensive experiments on newly translated VisDial datasets, we compare language-specific encoders (BERTimbau for Portuguese, BETO for Spanish) against multilingual BERT, achieving competitive performance with monolingual models while revealing distinct cross-modal attention patterns. Our mechanistic interpretability analysis demonstrates that despite different tokenization strategies and pretraining objectives, both approaches converge to similar attention distributions in deeper layers, with divergence decreasing from 0.000832 (Layer 0) to 0.000490 (Layer 11). We find that monolingual models exhibit holistic attention strategies while multilingual models show more selective, fine-grained visual grounding. These findings have important implications for developing inclusive vision-language technologies and understanding cross-lingual transfer in multimodal contexts.

Image segmentation is inherently challenging because it is often difficult to automatically identify the object of interest in an image. To mitigate this, human-provided markers can be incorporated into the segmentation process, greatly improving accuracy. However, human interaction is an expensive resource and methods that reduce effort in interactive segmentation are of great interest. In this work, we introduce a novel marker-based segmentation layer for deep neural networks, enabling end-to-end training of a marker creation network. Our training methodology includes a loss function with two main components: (i) segmentation loss using the new differentiable segmentation layer and (ii) a set of regularization functions that ensure the generated markers have the desired shape properties. We show that by using the proposed method, the network can automatically generate markers that achieve effective segmentation and have desirable shape characteristics. We validate our results in the training dataset and in five unseen datasets.

Tropical forests play a central role in the planet's ecosystem, making the conservation of these biomes a global priority. However, ongoing deforestation and degradation represent a significant threat to its existence, which requires effective monitoring and the proposition of actions to mitigate the damage caused by these processes. In this sense, several government and private monitoring programs have been developed, which apply image processing, machine learning and citizen science (CS). Mainly in the context of CS, an important project is ForestEyes, which segments remote sensing images to identify deforested areas and subsequently submit these segments for analysis by non-specialized volunteers. As segmentation using techniques based on superpixels proves to be a viable solution for this important task. Therefore, this paper presents an analysis of 22 segmentation methods based on superpixels applied to remote sensing images, aiming to identify which of them are most suitable for generating better segments for CS campaigns. The results revealed that seven of the segmentation methods outperformed the baseline method (SLIC) currently employed in the ForestEyes project, indicating an opportunity for improvement in this important phase of campaign development.

Few-shot semantic segmentation (FSS) aims to segment novel classes in query images given only a few annotated support examples. This task remains challenging due to the limited supervision available during adaptation. In this work, we propose the Saliency-guided Model with Cross and Channel Attention (SaMoCCA) for FSS, a novel architecture that enhances support guidance using class-agnostic saliency maps and performs dense alignment between query and support features through multi-layer cross-attention. Saliency cues are extracted using a pretrained model and fused with the support mask to highlight discriminative regions. The resulting attention-weighted features are refined hierarchically via multi-scale aggregation blocks, squeeze-and-excitation modules, and a final mask-feature mixer that restores spatial details using high-resolution encoder features. In summary, the major contributions of this work are the use of saliency maps to guide the support branch and the adoption of channel attention mechanism to adaptively recalibrate features. Our findings confirm that both the saliency-guided mask and the squeeze-and-excitation block contribute positively and complement each other. Extensive experiments on the PASCAL-5 ${ }^{i}$ and COCO-20 ${ }^{i}$ benchmarks demonstrate that the proposed approach achieves competitive results compared to recent state-of-the-art methods, while offering architectural simplicity, strong generalization to unseen classes, and also demonstrates robust cross-domain generalization from COCO-$20^{i}$ to PASCAL-5 ${ }^{i}$ without domain adaptation, highlighting the transferability of saliency-guided cross-attention.

The necessity of abundant annotated data and complex network architectures presents a significant challenge in deep-learning Salient Object Detection (deep SOD) and across the broader deep-learning landscape. This challenge is particularly acute in medical applications in developing countries with limited computational resources. Combining modern and classical techniques offers a path to maintaining competitive performance while enabling practical applications. Feature Learning from Image Markers (FLIM) methodology empowers experts to design convolutional encoders through user-drawn markers, with filters learned directly from these annotations. Recent findings demonstrate that coupling a FLIM encoder with an adaptive decoder creates a flyweight network suitable for SOD, requiring significantly fewer parameters than lightweight models and eliminating the need for backpropagation. Cellular Automata (CA) methods have proven successful in data-scarce scenarios but require proper initialization – typically through user input, priors, or randomness. We propose a practical intersection of these approaches: using FLIM networks to initialize CA states with expert knowledge without requiring user interaction for each image. By decoding features from each level of a FLIM network, we can initialize multiple CAs simultaneously, creating a multi-level framework. Our method leverages the hierarchical knowledge encoded across different network layers, merging multiple saliency maps into a high-quality final output that functions as a CA ensemble. Benchmarks across two challenging medical datasets demonstrate the competitiveness of our multi-level CA approach compared to established models in the deep SOD literature.

Oral Epithelial Dysplasia (OED) is a histopathological diagnosis for potentially malignant lesions with a variable risk of malignant transformation. The conventional grading of OED relies on morphological assessment, a process known for significant inter-observer variability, which limits its prognostic reliability. This study leverages computational pathology to address this challenge by employing a deep learning-based tool, Cellpose, for the precise segmentation of epithelial cell nuclei. The objective was to facilitate the objective extraction and analysis of key nuclear morphometric features, such as area, perimeter, and compactness, from digitized histological images of OED. We utilized 30 H&E-stained images from a spectrum of diagnostic categories, including normal oral mucosa, dysplasias of varying grades, and oral squamous cell carcinoma (OSCC). The results demonstrated that quantitative features, particularly the variance in nuclear area (anisonucleosis), serve as a robust differentiator between diagnostic grades, with the most significant changes observed in OSCC. By providing a reproducible and quantitative framework, this deep learning-driven approach represents a significant step towards a standardized diagnostic support system for OED, with future potential for accurately predicting malignant transformation risk.

Superpixel segmentation partitions multiple objects into disjoint parts so that their delineation can be accurately achieved by their grouping, and it has been used as an intermediary step for solving multiple problems. However, state-of-the-art algorithms face a significant challenge of effective and efficient segmentation irrespective of the problem’s domain (object and background characteristics, and user’s desires). In this work, we address such challenge by proposing several contributions. One of such is a novel superpixel segmentation framework, named Superpixels through Iterative CLEarcutting (SICLE), which generalizes two other contributions of this work. In SICLE, three independent steps are defined: (i) seed oversampling; (ii) superpixel generation using the Image Foresting Transform (IFT) framework; and (iii) seed removal. From (i), where a significantly high amount of seeds is selected, steps (ii) and (iii) are performed for generating superpixels from a refined seed set until achieving the desired number of superpixels. SICLE overcomes domain shifts primarily through steps (ii) and (iii), where the user may provide an objective function for optimization. Experimental results show that SICLE variants surpass several state-of-the-art algorithms concerning speed and accuracy for distinct domains while generating a series of segmentations in a single execution. Still, in SICLE, the contours from a preceding scale might not be present in the subsequent one leading to hierarchical violations. Thus, we studied eight possible cases when analyzing pairwise subsequent segmentations, and we conceived three measures for estimating the hierarchiness of a multiscale segmentation: (i) nestedness; (ii) inflation ratio; and (iii) refinement error. From our results, it is possible to verify if a multiscale is a hierarchy and, when it is not the case, to analyze and state the nature and extent of the hierarchical violations that prevent it from being hierarchical.

Superpixels through Iterative CLEarcutting (SICLE) is an efficient framework for image segmentation that operates on an iterative principle of seed removal to refine the final result. While effective for general purposes, its reliance on standard seed removal criteria limits its performance in specialized domains. This is particularly evident in biomedical analysis, where the goal is often to isolate a single object of interest using a low number of superpixels, a task where generic criteria often fail. To address this limitation, this work proposes two seed removal criteria: (i) a Position-Based Criterion that leverages prior anatomical knowledge to guide segmentation in medical images; (ii) a Color-Based Criterion specialized for identifying targets in pathological images based on their distinct color signature. By replacing generic heuristics with these domain-specific functions, we demonstrate that SICLE can be transformed into a more robust and specialized tool for targeted biomedical image analysis, significantly improving object delineation accuracy, including situations with a small quantity of superpixels.

Capturing relevant content from videos while preserving temporal coherence remains a central challenge in video skimming. The prevalence of redundant information often hinders the extraction of meaningful content, especially when the goal is to retain the central narrative of the video. While scene change detection can aid in segmenting video content, conventional methods often struggle with highly diverse and repetitive scenes due to their limited ability to model temporal dependencies and detect transitions effectively. To address these limitations, we propose the Memory-Augmented LSTM for Dynamic Video Summarization - MALSumm, a supervised architecture based on Extended Long Short-Term Memory (xLSTM) networks that enhances memory capacity through a dual-path design. This design integrates weighted memory to evaluate local and global information, allowing the model to preserve fine-grained details while maintaining overall temporal consistency, all within a low-complexity framework. Experimental Results validate the effectiveness of our approach, achieving an average $\text{F}$-score of $\text{4 9. 7}$ on the SumMe dataset and 62.1 on TVSum, outperforming recent supervised baselines. Additionally, when measuring alignment with human annotations, the model attains a Kendall's $\tau$ of $\text{0. 1 8 0}$ and Spearman's $\rho$ of 0.242, exceeding the scores reported for human agreement. These findings demonstrate that our method provides a competitive and lightweight solution for dynamic video summarization, effectively balancing accuracy and efficiency.

Numerous segmentation methods are able to produce several partitions of the same image by tuning a scale parameter. In such a series of multilevel segmentations, if every region at a given level is included in a single region of the segmentation at the next level, then the series is called a hierarchy. Hierarchies are often desired for multiscale image representation and analysis due to their mathematical properties, leading to accurate and efficient solutions. Although certain effective strategies may not produce a hierarchy, it is uncertain whether their multiscale output is close to be one. This work explores several cases when analyzing two consecutive segmentations, as full inflation and full merge, for instance. From those, we provide three measures for evaluating the hierarchiness between two subsequent partitions: (i) nestedness; (ii) refinement error; and (iii) inflation ration. Using our proposals in a in-sequence pairwise comparison, as shown by the experimental results, it is possible to verify whether a multiscale segmentation is a hierarchy and, if not, to analyze the nature and extent of the hierarchical errors that prevent it from becoming hierarchical.

This work addresses the importance of incorporating multi-scale information in image representation by proposing a novel approach utilizing hierarchical segmentation and graph neural networks (GNNs). The proposed model, named Hierarchical Image Graph with Scale Importance (HIGSI), leverages hierarchical segmentation to construct graphs that capture relationships between nodes across different scales. This multi-scale representation simultaneously captures intricate details and global context, leading to a richer understanding of image structure than traditional methods. Additionally, a novel Region Graph Readout (RGR) function is introduced to assess the significance of each scale within the graph representation. By combining this multi-scale representation and the RGR function, HIGSI achieves competitive performance on image classification tasks, using smaller graphs or having fewer parameters than existing methods. This work also presents a comparative study with another hierarchical approach and an assessment of HIGSI’s components to investigate its decision-making process and its components’ contribution to the overall performance.