Thiago Lopes Trugillo da Silveira
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Thiago Lopes Trugillo da Silveira has a total of 79 co-authors. Here are the top co-authors by number of shared publications:
- Fabio Mariano Bayer
Universidade Federal de Santa Maria
12 shared publications - Renato José de Sobral Cintra
Universidade Federal de Pernambuco
11 shared publications - Alice Kozakevicius
Universidade Federal de Santa Maria
5 shared publications - Arjuna Madanayake
Florida International University, Miami, USA
4 shared publications Claudio R. Jung
UFRGS
4 shared publications- Claudio Rosito Jung
Institute of Informatics, Federal University of Rio Grande do Sul, Porto Alegre, Brazil
4 shared publications
Latest Publications
- Open Access
- 30 January 2026
- Published by Universidade Federal do Rio Grande do Sul in Revista de Informática Teórica e Aplicada
- Vol. 33 (1) , 61-68
Evaluating Contactless Fingerprint Segmentation for Interoperable Biometric Identification Systems
Fingerprint recognition is a popular and cost-effective biometric technology. Most existing systems require specialized hardware to capture fingerprints, but contactless fingerprint capture can be performed using smartphone images. This method reduces costs and enhances hygiene and usability. A new image processing workflow is needed to facilitate contactless fingerprint capture, with segmentation a critical step. This study explores the feasibility of using smartphone images for this purpose. It evaluates four deep learning models - SSD MobileNetV2, SSD MobileNetV2 FPN Lite, Mask R-CNN, and U-Net - using the ISPFDv2 dataset. Results show that Mask R-CNN performed best in segmenting fingertip regions, while SSD MobileNetV2 had the highest recognition accuracy against traditional fingerprint databases. These findings demonstrate the potential of deep learning for effective contactless fingerprint recognition using smartphones.
- 30 January 2026
- Published by Universidade Federal do Rio Grande do Sul in Revista de Informática Teórica e Aplicada
- Vol. 33 (1) , 61-68
Open AccessEvaluating Contactless Fingerprint Segmentation for Interoperable Biometric Identification Systems
Fingerprint recognition is a popular and cost-effective biometric technology. Most existing systems require specialized hardware to capture fingerprints, but contactless fingerprint capture can be performed using smartphone images. This method reduces costs and enhances hygiene and usability. A new image processing workflow is needed to facilitate contactless fingerprint capture, with segmentation a critical step. This study explores the feasibility of using smartphone images for this purpose. It evaluates four deep learning models - SSD MobileNetV2, SSD MobileNetV2 FPN Lite, Mask R-CNN, and U-Net - using the ISPFDv2 dataset. Results show that Mask R-CNN performed best in segmenting fingertip regions, while SSD MobileNetV2 had the highest recognition accuracy against traditional fingerprint databases. These findings demonstrate the potential of deep learning for effective contactless fingerprint recognition using smartphones.
- Open Access
- 30 January 2026
- Published by Universidade Federal do Rio Grande do Sul in Revista de Informática Teórica e Aplicada
- Vol. 33 (1) , 28-40
Do Deep Learning Models Generalize Facial Emotion Recognition in Different Age Groups?
The prevalence of images of adults in facial emotion recognition (FER) databases is an evident bias in emotion classification. Although there are sets of FER images exclusively of children or adults aged 60 and over, the capture scenario is generally controlled and has staged emotions. Many studies in this area focus on increasing task performance by using a single data collection for training and testing; however, due to the diversity of individuals portrayed and photographs collected, this strategy generally fails when applied to new images or in real-world circumstances. In this paper, we assess the generalization potential of deep learning models that have been pre-trained in general-purpose databases such as ImageNet and refined using the databases CK+, DEFSS, FACES (60+), MUG, and NIMH-ChEFS with cross-databases in tests. Through this study, which assesses six different datasets involving three distinct age groups through cross-database evaluation using a clear experimental protocol, we were able to verify that models fine-tuned with data sets composed exclusively of children's images have accuracy above 80%, reaching 88% when used to predict images of children; models fine-tuned with images of young adults are more accurate in classifying young adults than children or adults aged 60 and over, and the dataset of adults aged 60 and over in FACES dataset demonstrated accuracy above 80% in predicting images from other datasets and 96% in predicting images within its own test set.
- 30 January 2026
- Published by Universidade Federal do Rio Grande do Sul in Revista de Informática Teórica e Aplicada
- Vol. 33 (1) , 28-40
Open AccessDo Deep Learning Models Generalize Facial Emotion Recognition in Different Age Groups?
The prevalence of images of adults in facial emotion recognition (FER) databases is an evident bias in emotion classification. Although there are sets of FER images exclusively of children or adults aged 60 and over, the capture scenario is generally controlled and has staged emotions. Many studies in this area focus on increasing task performance by using a single data collection for training and testing; however, due to the diversity of individuals portrayed and photographs collected, this strategy generally fails when applied to new images or in real-world circumstances. In this paper, we assess the generalization potential of deep learning models that have been pre-trained in general-purpose databases such as ImageNet and refined using the databases CK+, DEFSS, FACES (60+), MUG, and NIMH-ChEFS with cross-databases in tests. Through this study, which assesses six different datasets involving three distinct age groups through cross-database evaluation using a clear experimental protocol, we were able to verify that models fine-tuned with data sets composed exclusively of children's images have accuracy above 80%, reaching 88% when used to predict images of children; models fine-tuned with images of young adults are more accurate in classifying young adults than children or adults aged 60 and over, and the dataset of adults aged 60 and over in FACES dataset demonstrated accuracy above 80% in predicting images from other datasets and 96% in predicting images within its own test set.
- Open Access
- 20 January 2026
- Published by Sociedade Brasileira de Computação in Journal of the Brazilian Computer Society
- Vol. 32 (1) , 1471-1485
Adapting Convolutions for Effective Omnidirectional Image Processing
Omnidirectional images present unique challenges for traditional convolutional neural networks (CNNs) due to the non-uniform sampling inherent in the equirectangular projection (ERP). This projection introduces distortions, especially near the poles, and conventional fixed-size kernels in planar CNNs are not designed to handle these distortions effectively. To address this issue, we previously introduced the Spherically-Weighted Horizontally Dilated Convolutions (SWHDC) module which adjusts dilated convolutions by applying appropriate weights to account for the varying sampling density across ERP image rows. In this extended work, we provide a more comprehensive evaluation of the SWHDC module by benchmarking it against several state-of-the-art methods on the 3D object classification task. Additionally, we integrate SWHDCs into different backbones to further investigate their effectiveness in tackling the gravity alignment problem. Experimental results confirm that our approach not only improves classification accuracy but also enhances gravity alignment performances without increasing the amount of trainable parameters of the baseline backbones. These findings further establish SWHDC as a robust alternative for processing omnidirectional images in different visual computing applications.
- 20 January 2026
- Published by Sociedade Brasileira de Computação in Journal of the Brazilian Computer Society
- Vol. 32 (1) , 1471-1485
Open AccessAdapting Convolutions for Effective Omnidirectional Image Processing
Omnidirectional images present unique challenges for traditional convolutional neural networks (CNNs) due to the non-uniform sampling inherent in the equirectangular projection (ERP). This projection introduces distortions, especially near the poles, and conventional fixed-size kernels in planar CNNs are not designed to handle these distortions effectively. To address this issue, we previously introduced the Spherically-Weighted Horizontally Dilated Convolutions (SWHDC) module which adjusts dilated convolutions by applying appropriate weights to account for the varying sampling density across ERP image rows. In this extended work, we provide a more comprehensive evaluation of the SWHDC module by benchmarking it against several state-of-the-art methods on the 3D object classification task. Additionally, we integrate SWHDCs into different backbones to further investigate their effectiveness in tackling the gravity alignment problem. Experimental results confirm that our approach not only improves classification accuracy but also enhances gravity alignment performances without increasing the amount of trainable parameters of the baseline backbones. These findings further establish SWHDC as a robust alternative for processing omnidirectional images in different visual computing applications.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 130575-130588
Geometry-Aware Transform Pruning for Omnidirectional Image Compression
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 130575-130588
Open AccessGeometry-Aware Transform Pruning for Omnidirectional Image Compression
- 29 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Design & Test
- Vol. 43 (2) , 54-63
Exploring Asymmetric Autoencoder Architectures for Computationally Efficient Neural Image Compression
Editor’s notes: This article explores asymmetric autoencoder architectures for neural image compression, reducing decoder complexity to improve efficiency. By pruning decoder layers, the proposed models achieve up to 46% faster CPU and 18% faster GPU decompression with only minor bitrate increases, enabling more practical neural image compression deployment in resource-constrained environments. —José Augusto Nacif, Federal University of Viçosa, Brazil —Omar Paranaiba Vilela Neto, Federal University of Minas Gerais (UFMG), Brazil
- 29 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Design & Test
- Vol. 43 (2) , 54-63
Exploring Asymmetric Autoencoder Architectures for Computationally Efficient Neural Image Compression
Editor’s notes: This article explores asymmetric autoencoder architectures for neural image compression, reducing decoder complexity to improve efficiency. By pruning decoder layers, the proposed models achieve up to 46% faster CPU and 18% faster GPU decompression with only minor bitrate increases, enabling more practical neural image compression deployment in resource-constrained environments. —José Augusto Nacif, Federal University of Viçosa, Brazil —Omar Paranaiba Vilela Neto, Federal University of Minas Gerais (UFMG), Brazil
- Open Access
- 9 June 2025
- Published by Sociedade Brasileira de Computação
- p. 713-724
Complexity-Reduced End-to-End Fetal ECG Signal Recovery and QRS Complex Detection
Non-invasive electrocardiogram (niECG) enables pregnancy monitoring by assessing the mother’s and fetus’ health through maternal abdominal signal acquisition. However, isolating fetal ECG (fECG) is challenging due to low signal-to-noise ratio and time and frequency overlap with maternal cardiac activity. Prior studies focused on fetal R-peak detection and more recent works target full fECG waveform noise-free reconstruction, but at a high computational cost, precluding real-world applications. This paper explores different ways of reducing the complexity of state-of-the-art deep learning-based method for fECG recovery and fetal QRS complex localization. Results indicate that one encoder-decoder block pair can be removed without significantly impacting the metrics. While the other complexity-reduction options yield slightly lower metrics compared to the baseline and related works, they can be fine-tuned and adapted based on the specific requirements and objectives of the application.
- 9 June 2025
- Published by Sociedade Brasileira de Computação
- p. 713-724
Open AccessComplexity-Reduced End-to-End Fetal ECG Signal Recovery and QRS Complex Detection
Non-invasive electrocardiogram (niECG) enables pregnancy monitoring by assessing the mother’s and fetus’ health through maternal abdominal signal acquisition. However, isolating fetal ECG (fECG) is challenging due to low signal-to-noise ratio and time and frequency overlap with maternal cardiac activity. Prior studies focused on fetal R-peak detection and more recent works target full fECG waveform noise-free reconstruction, but at a high computational cost, precluding real-world applications. This paper explores different ways of reducing the complexity of state-of-the-art deep learning-based method for fECG recovery and fetal QRS complex localization. Results indicate that one encoder-decoder block pair can be removed without significantly impacting the metrics. While the other complexity-reduction options yield slightly lower metrics compared to the baseline and related works, they can be fine-tuned and adapted based on the specific requirements and objectives of the application.
- Open Access
- 9 June 2025
- Published by Sociedade Brasileira de Computação
- p. 115-120
Deep learning applied to non-invasive ECG pregnancy monitoring: fetal ECG recovery and QRS complex detection
Non-invasive electrocardiography (niECG) is an option to mitigate the limitations of the current pregnancy monitoring methods. Still, one of the main challenges in its use is the extraction of the fetal signal (fECG) from the measured abdominal signal (aECG) due to the low signal-to-noise ratio and the overlap of the maternal and fetal QRS complexes in time and frequency domains. We present two encoder-decoder deep learning models that use regions of interest (RoI) to perform simultaneous fetal QRS complex (fQRS) detection and fECG recovery. We show that such a RoI-based end-to-end approach leads to robust state-of-the-art results in inter-subject and cross-dataset evaluations.
- 9 June 2025
- Published by Sociedade Brasileira de Computação
- p. 115-120
Open AccessDeep learning applied to non-invasive ECG pregnancy monitoring: fetal ECG recovery and QRS complex detection
Non-invasive electrocardiography (niECG) is an option to mitigate the limitations of the current pregnancy monitoring methods. Still, one of the main challenges in its use is the extraction of the fetal signal (fECG) from the measured abdominal signal (aECG) due to the low signal-to-noise ratio and the overlap of the maternal and fetal QRS complexes in time and frequency domains. We present two encoder-decoder deep learning models that use regions of interest (RoI) to perform simultaneous fetal QRS complex (fQRS) detection and fECG recovery. We show that such a RoI-based end-to-end approach leads to robust state-of-the-art results in inter-subject and cross-dataset evaluations.
- 28 April 2025
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 84 (34) , 43351-43368
Superpixel-driven 360$^\circ $ image compression- 28 April 2025
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 84 (34) , 43351-43368
Superpixel-driven 360$^\circ $ image compression- Open Access
- 19 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 105119-105130
An Approach Based on LiDAR and Spherical Images for Automated Vegetation Inspection in Urban Power Distribution Lines
Preventive tree pruning is an efficient way to reduce the risks of outages in the electrical network, such as cable rupture, short circuits, and fires, which may occur due to interferences between vegetation and urban electrical distribution lines. These interferences are listed as one of the main causes of power supply interruptions by distribution companies and, consequently, generate significant financial losses due to the decline in quality indicators and the need for maintenance. Vegetation management inspection is essential to maintain the reliability and continuity of the power distribution, but manual monitoring is expensive, time-consuming, and subjective. This paper proposes a new method based on Deep Learning (DL) and high-resolution sensors for automated vegetation inspection in urban power distribution lines. The method combines Light Detection and Ranging (LiDAR) and spherical images and overcomes the limitations of aerial images using vehicle-borne sensors. The dataset was obtained using the proposed capture system in four Brazilian cities. The results demonstrated the suitability of the proposed method in an innovative way through a proof of concept. Among the main contributions, the functional evaluation of the system for Artificial Intelligence-based real-time vegetation inspection in power line distribution is highlighted. The proposed method achieved high rates of interference identification using a web platform and mobile application (achieving a precision of over 94%). The results demonstrate the potential of the proposed approach for automated vegetation inspection in real-world scenarios and open up opportunities for the deployment and evaluation of large-scale applications.
- 19 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 105119-105130
Open AccessAn Approach Based on LiDAR and Spherical Images for Automated Vegetation Inspection in Urban Power Distribution Lines
Preventive tree pruning is an efficient way to reduce the risks of outages in the electrical network, such as cable rupture, short circuits, and fires, which may occur due to interferences between vegetation and urban electrical distribution lines. These interferences are listed as one of the main causes of power supply interruptions by distribution companies and, consequently, generate significant financial losses due to the decline in quality indicators and the need for maintenance. Vegetation management inspection is essential to maintain the reliability and continuity of the power distribution, but manual monitoring is expensive, time-consuming, and subjective. This paper proposes a new method based on Deep Learning (DL) and high-resolution sensors for automated vegetation inspection in urban power distribution lines. The method combines Light Detection and Ranging (LiDAR) and spherical images and overcomes the limitations of aerial images using vehicle-borne sensors. The dataset was obtained using the proposed capture system in four Brazilian cities. The results demonstrated the suitability of the proposed method in an innovative way through a proof of concept. Among the main contributions, the functional evaluation of the system for Artificial Intelligence-based real-time vegetation inspection in power line distribution is highlighted. The proposed method achieved high rates of interference identification using a web platform and mobile application (achieving a precision of over 94%). The results demonstrate the potential of the proposed approach for automated vegetation inspection in real-world scenarios and open up opportunities for the deployment and evaluation of large-scale applications.
- 29 April 2024
- Published by Springer Nature in The Visual Computer
- Vol. 40 (6) , 3961-3970
Triangular matrix-based lossless compression algorithm for 3D mesh connectivity
- 29 April 2024
- Published by Springer Nature in The Visual Computer
- Vol. 40 (6) , 3961-3970
Triangular matrix-based lossless compression algorithm for 3D mesh connectivity