Melanoma is the deadliest skin cancer, causing over 90% of skin tumor deaths. Early detection ensures survival above 95%, but below 15% in late stages. We propose the HEVA architecture, which integrates ensemble deep learning with ABCD descriptors and a SegFormer-B5 model, for accurate lesion masks. HEVA achieved 93.2% accuracy and 84.2% F1-score on HAM-10000, and 87.7% accuracy with 82.01% macro F1 on ISIC 2018 (1,512 images). SegFormer achieved 82.6% mean IoU and 91.2% accuracy in segmentation.
Coffee is one of the main agricultural products exported by Brazil, with Minas Gerais standing out as the largest producer and responsible for a significant portion of the country's production. Due to its importance, this work investigates various Convolutional Neural Networks (CNNs) for mapping coffee plantation areas in the municipality of Muzambinho, MG. Using satellite images, we evaluated four semantic segmentation network architectures: FCN, MA-NET, SegNet, and DeepLabV3. Experiments using a dataset with 100 satellite images demonstrated that the FCN and MA-NET networks achieve better performance in identifying crops than SegNet and DeepLabV3. The latter, however, presented generalization challenges in the test set, indicating sensitivity to variations in lighting and shading. The study highlights the potential of artificial intelligence combined with remote sensing to improve agricultural management, and the need for more robust and diversified datasets to optimize the generalization and robustness of the models in real conditions.
Accurate segmentation of spinal structures in magnetic resonance imaging (MRI) is an important step for supporting the diagnosis of several diseases and related conditions. In this work, we present a systematic evaluation of three deep learning architectures, U-Net, Feature Pyramid Network (FPN), and SegFormer, combined with ResNet-50 and EfficientNet-B2 encoders for the semantic segmentation of the cervical spine (C1–C7) in MRI volumes from the VerSe 2020 dataset. The proposed pipeline includes dataset preparation, model training using Dice loss, Adam optimizer, and early stop strategy, and evaluation with standard metrics such as Intersection over Union (IoU), F1-Score, and accuracy. Results show that FPN with ResNet-50 achieved the best overall performance, reaching an IoU of 0.6696 and an F1-score of 0.8021, while EfficientNet-B2 provided more consistent results across different architectures. Data augmentation showed limited impact, with gains restricted to a few specific configurations. Qualitative analysis through 3D surface reconstruction further highlighted the limitations of 2D slice-based segmentation, particularly at the extremities of the vertebrae, suggesting the need for volumetric approaches. The contributions of this work include the development of a reproducible pipeline for spinal segmentation, a comparative evaluation of convolutional and Transformer-based models. Qualitative analysis suggests potential benefits in exploring 3D approaches in future work.
Automated quality control in industrial production has the potential to reduce errors and provide real-time information. However, the inspection of secondary packaging, such as counting boxes in crates, still represents a challenge, since manual methods are slow and error-prone, while automatic methods are limited by the scarcity of domain-specific datasets and the high cost of annotation. This paper proposes an efficient and low-cost two-stage object detection workflow for automatic box counting. The central novelty lies in the integration of the Segment Anything Model (SAM) to accelerate the creation of a high-quality dataset from production line videos, making model specialization economically feasible. Initially, a generalist model is trained on heterogeneous datasets to capture general visual features. Then, the model is fine-tuned with the domain-specific dataset of approximately 750 images. Experiments with YOLOv11x and Faster R-CNN achieved mAP@0.5 above 98%, with YOLOv11x showing higher accuracy and faster inference. These results demonstrate the efficiency of the proposed approach, establishing it as a replicable and low-cost solution for monitoring secondary packaging in industrial environments.
Air writing is a form of human–computer interaction that enables users to write in the air through different technological approaches, such as radio waves, wearable sensors, dedicated devices, and computer vision. In this work, we present a hybrid solution that combines the use of a luminous-tip pen with computer vision techniques to support pen detection and the air-writing process. The pen was designed and fabricated with a 3D printer and integrates an ESP32 microcontroller board, which provides Bluetooth connectivity for real-time communication with the computer. In addition to tracking movements for writing, the device can transmit specific commands to the system, such as starting or stopping the writing process and changing colors. This approach bridges the gap between wearable devices and vision-based methods, offering a practical and accessible alternative for air-writing applications. The system integrates vision–language models capable of recognizing both words and images. To determine which models were best suited for our system, we evaluated the performance of several candidates. Five participants contributed to the creation of the dataset. For drawing recognition, each participant produced images in the following classes: Tree, Moon, Cat, Heart, and Pen. For word recognition, the dataset included the following eight words: Purple, Window, Jungle, Pillow, Team, Doctor, Words, and Science. The results showed that the best-performing models achieved 76.00% accuracy with Kosmos-2 for image detection and 97.39% accuracy with Gemini 2.5 Flash for word detection, demonstrating satisfactory outcomes for the system implementation. The proposed system demonstrates the potential of combining electronic devices and advanced machine learning techniques for air-writing, with applications in education and assistive technologies.
Big Data systems are expanding rapidly, and frameworks like Hadoop and Spark are now central to that growth. Yet this expansion also raises new security challenges, particularly regarding how vulnerabilities are identified, assessed, and mitigated in complex, distributed environments. While the existing literature predominantly focuses on defensive mechanisms, systematic evidence on offensive approaches such as penetration testing remains limited. This paper presents a systematic literature review of security assessment practices for Hadoop and Spark between 2015 and 2025, with an emphasis on penetration-testing techniques in Big Data frameworks. Our search across major digital libraries retrieved 1578 records. After title and abstract screening, 46 articles were selected. Of these, 20 were considered relevant and read in full; among these, only four explicitly applied penetration testing to Hadoop or Spark deployments. The review reveals a shortage of realistic testing environments and standardized metrics for evaluating the effectiveness of mitigation strategies, alongside a strong reliance on generic security tools rather than specialized offensive frameworks. These findings underscore the need for dedicated, framework-aware penetration-testing solutions for Hadoop and Spark, reproducible testing protocols, and tighter integration between offensive assessments and proactive mitigation practices.
Deep learning-based monocular depth estimation has achieved significant advancements on urban benchmarks, but its embedded application remains limited by efficiency constraints. Vision Transformers (ViTs) and Foundation Models (FMs) show promising zero-shot generalization capabilities, yet their adaptation to resource-constrained hardware requires careful study. In this work, we investigate the development of the DepthAnything model on an NVIDIA Jetson Orin, analyzing the trade-off between accuracy and inference speed for different backbones (ViT-S, ViT-B, and ViT-L). We report quantitative metrics including AbsRel, δ1, RMSE, and FPS on the KITTI dataset, along with qualitative results. Our experiments show that the ViT-S backbone offers the best balance of accuracy and real-time performance (44 FPS), whereas ViT-B suffers from degradation and ViT-L exhibits significant instability due to optimization artifacts. These findings highlight the viability of compact backbones for embedded visual perception and suggest future optimizations, such as quantization-aware training and pruning, in larger architectures.
Breast cancer detection is a global health priority. While traditional methods have limitations, infrared thermography offers a promising, non-invasive alternative by detecting subtle thermal changes that can indicate tumors. This paper assessed five pre-trained Convolutional Neural Networks (CNNs) for breast cancer detection using DMR-IR thermal images, employing a 5-fold cross-validation. Among the tested models, ResNet50 achieved the best overall performance, with the highest average accuracy (92.79%), precision (95.00%), specificity (98.67%), sensitivity (72.00%), and F1-score (79.43%). The model was trained using raw thermal images from three anatomical views (frontal, lateral 90°, and lateral 45°), totaling five images per patient, an approach still uncommon in the literature. These results highlight ResNet50's strong potential for reliable and clinically applicable breast cancer detection using thermography.
Wildfires pose serious threats to ecosystems and human safety, requiring accurate monitoring systems. This study proposes a Federated Learning (FL) approach with Convolutional Neural Networks (CNNs) for wildfire detection using two heterogeneous image datasets, keeping the data locally on each client. The federated setup simulates non-Independent and identically distributed (IID) conditions, where each client trains locally and updates are aggregated its weights to a remote server. To ensure effective performance, hyperparameter optimization for each architecture was conducted using the Tree of Parzen Estimators (TPE), allowing efficient exploration of the best training configurations. Results demonstrate that FL can handle data heterogeneity while preserving privacy, with deeper CNN architectures achieving superior performance. The findings highlight the feasibility of FL for wildfire surveillance and the ability of optimized CNNs to generalize effectively across diverse environmental conditions, supporting collaborative model training without sharing raw data.
This study evaluates the performance of three low-cost 3D scanning techniques: structured light sensors (using a Kinect and a CR-Scan Ferret) and photogrammetry, to capture 3D-printed physical models. The research compares the resulting 3D models with the original digital models to assess each technique's accuracy and the effectiveness of low-cost methods. The study’s workflow involved printing four 3D models, scanning them, and comparing the models using similarity metrics like Hausdorff distance, RMSE, and Chamfer distance. The CR-Scan Ferret generally outperformed the other methods, particularly in capturing detail. However, it sometimes produced gaps that required filling. Photogrammetry was a viable alternative, excelling at preserving object shapes without creating gaps. The Kinect, while accessible, consistently delivered the lowest performance, struggling to reproduce object details and forms accurately.
Visual Language Models (VLMs) combine natural language processing and computer vision to interpret multimodal data, such as images and text, showing great potential in image classification applications. This paper investigates the integration of Active Learning (AL) and pseudo-labeling techniques with VLMs to improve image classification in various domains. To achieve this, five AL strategies (Random Sampling, Uncertainty Sampling, Margin Sampling, Entropy Sampling, and Query-by-Committee) and three pseudo-labeling approaches (Direct, Confidence Threshold, and Feature Similarity) were evaluated iteratively. The results demonstrate that the combination of active learning and pseudo-labeling can achieve promising results, in addition to full class coverage in a few iterations. We conclude that the integration of AL with feature similarity-based pseudo-labeling offers a robust and efficient solution for image classification in limited-data scenarios, promoting high accuracy, class representativeness, and the reduction of propagation errors, with potential for applications in critical domains like healthcare and industry.
The increasing adoption of Blended Learning (BL) models in Brazilian education, particularly after the COVID-19 pandemic, has led to diverse approaches and inconsistent implementations. Motivated by this scenario, this study aimed to analyze how BL has been conceptualized and applied across different contexts. A Systematic Literature Review (SLR) of studies published between 2013 and 2022 was conducted, followed by a hierarchical clustering process to identify conceptual similarities. As a result, five distinct BL models were identified: Integrated, Connected, Intra-school, Team-Based, and Highly Tutored. These models were synthesized into a Conceptual Structure, which was subsequently organized into conceptual frameworks to support educators’ decision-making. To explore the applicability of the proposed structure, three of the identified models were implemented in 8th-grade classrooms in a Brazilian public school in Minas Gerais. Each group experienced a different BL model over a five-week period, and students were evaluated using performance measures, Likert-scale self-assessments, and statistical analyses. The results indicate higher levels of student engagement and academic performance in the Team-Based model when compared to the other two models, while the Intra-school model, although supported by existing school infrastructure, showed limitations in fostering student autonomy and participation. The findings indicate that there is no single hybrid learning model applicable to all contexts. Its effectiveness depends on pedagogical alignment, technological infrastructure, and students’ profiles. The study provides a practical conceptual framework and empirically grounded references to guide teachers and educational managers in implementing hybrid learning in Brazil.
This work addresses the challenge of classifying genosensor images obtained by scanning electron microscopy (SEM) for SARS-CoV-2 diagnosis. Due to the specialized nature of the data, image datasets are inherently small, which limits the effectiveness of deep learning models. To overcome this limitation, we investigated the impact of seven data augmentation strategies - Flipping, Shift, Random Erasing, AugMix, AutoAugment, RandAugment and TrivialAugment - on the performance of four Convolutional Neural Networks (CNNs): ResNet50, ResNet101, DenseNet121, and ConvNeXt-Tiny. Our analyses demonstrated that data augmentation is crucial for the success of the task, achieving a remarkable accuracy of 97.85%, surpassing both traditional feature extraction methods and previously tested deep learning approaches. Furthermore, our findings indicate that the choice of augmentation technique is critical, with aggressive methods proving detrimental for highly specialized microscopy data.
In recent years, the integration of large language models and knowledge graphs has increasingly attracted the attention of both researchers and practitioners. While existing reviews provide valuable and helpful insights for better understanding the research field, they lack a quantitative perspective. To address this limitation, a bibliometric analysis was performed in this study to fill the gap. In particular, sourcing data from the Web of Science and Scopus databases, our analysis focuses on: (1) publication and citation trends, (2) the most productive countries and authors, (3) the most influential sources and scientific papers, (4) the main research tasks particularly in KG-enhanced LLMs and LLM-augmented KGs, (5) the most frequently employed techniques and their evolution over time. In addition to providing a comprehensive overview of the most active and recent areas of research, this study identifies several emerging themes, namely: “Knowledge-aware Prompt Engineering”, “Retrieval-Augmented Generation for LLM–KG Integration”, “Knowledge Graph-based Hallucination Detection” and the convergence of “Multimodal Knowledge Graphs and Multimodal Large Language Models”. These insights contribute to a clearer understanding of the field's development and highlight emerging research trends grounded in bibliometric evidence.
This work presents a study on speckle noise filtering in SAR images using convolutional neural networks. The objective is to develop an accessible and efficient solution to improve the quality of images degraded by this type of noise, facilitating their analysis and interpretation. The proposed methodology involves the using synthetic SAR-like data, the training of a convolutional neural network, and the evaluation of results using traditional metrics such as PSNR and SSIM. Additionally, stochastic distances were integrated into the loss function of the convolutional neural network, enabling a more detailed analysis of the preservation of the statistical properties of the filtered images. The results indicate that the model effectively reduces speckle noise while preserving the structural details of the image. Furthermore, comparisons with recent studies on the topic showed that the proposed solution achieves competitive performance.
Colorectal cancer (CRC) is one of the highest incident cancers in the world. The late-stage diagnosis plays a pivotal role in the mortality rate, making CRC the second leading cause of cancer-related deaths. Its diagnosis is based on the analysis of histological images acquired from a biopsy, a time-consuming and prone to errors task. Over the years, many deep learning and computer vision approaches have been proposed to automatize such a task, reducing the need for human specialists. To contribute to this area of research, we proposed an ensemble that combines a Lightweight CNN and handcrafted color texture features commonly used in literature. We investigated how different texture methods impact the performance of the ensemble in an important multi-class problem composed of eight types of tissues. Our ensemble obtained 99.63% accuracy, surpassing state of the art methods and deeper CNNs, and 99.63% F1 score, showing a good balance between precision and recall.
Early detection of skin cancer is vital for effective treatment and improving patient recovery. In recent years, a growing number of computer vision studies have been developed to aid in diagnosis, drawing significant attention from researchers. However, challenges still persist, such as data imbalance and the lack of comprehensive datasets. Additionally, limited research has focused on how variations in skin tone across different populations affect the performance of models in skin lesion classification. This study seeks to create a more effective approach to address data biases in lesion classification across diverse skin tones. We initially explored several data augmentation techniques, employing traditional feature extractors for image analysis. For classification, models such as k-Nearest Neighbor, Random Forest, and Support Vector Machine were used. This study focused on two well-known and publicly available skin lesion datasets: HAM10000 and PAD-UFES-20, both of which have significant class imbalances. Further experiments were conducted to assess potential biases, with the Individual Typology Angle (ITA) metric applied to evaluate the skin tone distribution within the datasets.
This study presents a novel approach to texture analysis that combines Complex Network Texture Descriptors (CNTD) with the Local Binary Pattern (LBP) technique. By generating "pattern images" through the application of LBP to the original texture image, the method introduces new sources of information that are further explored via CNTD method. We employed Particle Swarm Optimization (PSO) to investigate various combinations of these LBP patterns and descriptors and compared our results across three benchmark datasets. The proposed approach achieved accuracy rates of 99.54%, 99.88%, and 97.86% on the Vistex, Brodatz, and USPTex databases, respectively. These results highlight the effectiveness of the hybrid strategy in producing a highly discriminative feature vector for robust texture classification.
In agricultural robotics and orchard automation, tasks such as fruit detection, tracking, and spatial localization are essential for applications like yield prediction and harvesting. However, these tasks are challenging due to the similar appearance of fruits, occlusions, and the inherent difficulties of field robotics, including uncontrolled lighting conditions and the variability in orchard environments. This work leverages CoTracker, a transformer-based model for point tracking using cross-track/cross-time attention, to simultaneously perform multiple fruit tracking, 3-D fruit localization, and camera pose estimation. The proposed approach demonstrates promising results in fruit counting, tracking, and scene reconstruction, highlighting its potential in agricultural automation.
Although the rapid advancement of Large Language Models (LLMs) has significantly transformed software development, enabling the automation of tasks such as code generation, testing, and documentation, concerns remain regarding the quality, efficiency, and security of automatically generated solutions, especially in relation to computational resource usage such as memory, CPU, and execution time. This scenario is intensified by the growing demand for digital sustainability and the need to minimize operational costs. This work investigates the efficiency and correctness of code generated by the LLMs Llama-3.3-70B Versatile and Gemma2-9B Instruct when solving Rosetta Code tasks in Python, comparing them with human implementations. The analyses address multiple aspects, including code correctness, clarity, memory consumption, CPU usage, and execution time, thus revealing the limitations and strengths of automatic solutions compared to human-developed code. The results suggest that, although LLMs are capable of quickly generating functional solutions, they still lack the algorithmic optimizations necessary to match the efficiency of manually written code, particularly regarding resource utilization. However, for routine or well-known tasks, automatic solutions can be highly competitive. This study highlights the importance of multidimensional evaluations and demonstrates that, if combined with human review and good engineering practices, the careful adoption of LLMs can enhance software development processes.
