Blog /

RESEARCH

Research
2025

Open-set Face Recognition

This research tackles the significant challenges of open-set face recognition for watchlist applications, focusing on identifying known individuals while reliably rejecting unknown subjects unseen during training. The core problem is that standard neural networks are biased towards known data, struggling to distinguish unknowns and operate effectively with limited samples per identity. Practical deployment also faces hurdles from the computational expense of retraining deep networks for gallery updates and challenging uncontrolled surveillance conditions.

Motivated by the need for robust and agile open-set systems that minimize false positives, this work introduces compact adapter networks for rapid domain adaptation and novel discriminative loss functions like Maximal Entropy Loss (MEL), Objectosphere Loss, and Axial Sphere Loss (ASL). These methods, leveraging real or synthesized negative samples, aim to enhance class separation and minimize the open-space risk by pushing non-gallery samples towards the feature space origin, demonstrating their effectiveness on datasets such as LFW, IJB-C, and UCCS.

References:
• Vareto, R. H., Linghu, Y., Boult, T. E., Schwartz, W. R., & Günther, M. (2024). Open-set face recognition with maximal entropy and Objectosphere loss. Image and Vision Computing, 141, 104862. https://doi.org/10.1016/j.imavis.2023.104862
• Kasıma, F., Boult, T. E., Mora, R., Biesseck, B., Ribeiro, R., Schlueter, J., Repák, T., Vareto, R. H., Menotti, D., Schwartz, W. R., & Günther, M. (2024). Watchlist Challenge: 3rd Open-set Face Detection and Identification.
• Vareto, R. H., & Schwartz, W. R. (2025). AXIAL SPHERE LOSS: Encouraging open-space risk minimization in face identification tasks.

2025

Face Anonymization Protection

This research addresses a critical and previously overlooked security flaw in state-of-the-art reversible face anonymization systems, which are designed to balance privacy and data utility but are not as secure as presumed. The core problem is that these systems leak original identity information in intermediate representations, enabling a novel attack to bypass password protection and recover identities at a high rate, reaching 70.2% success on the VggFace2 dataset against a leading method like RiDDLE.
 
Motivated by the urgent need to ensure genuine privacy and fortify these methods against such vulnerabilities in compliance with data protection regulations, this work first unveils the attack and then proposes a Secure Image Embedding Loss. This new loss function effectively enforces the disassociation of identity from intermediate embeddings, thereby reducing the attack's success rate to a mere 0.8% on VggFace2 and demonstrating robust cross-dataset performance.
 
 
References:
• Fortifying Reversible Face Anonymization: A Secure Image Embedding Loss to Prevent Identity Leakage.
2025

Analysis of Seismic Data

This research encompasses a systematic literature review on deep learning for seismic image segmentation, with a particular focus on facies segmentation, and simultaneously proposes a principled benchmark for lithofacies segmentation. The key problems addressed are the time-consuming, expensive, and subjective nature of manual seismic interpretation, compounded by the scarcity of publicly available annotated datasets and a critical lack of standardized evaluation protocols in the field.
 
Motivated by the need to enhance efficiency, reduce costs, and establish a framework for rigorous, comparable research in geosciences, this work systematically analyzes current deep learning methods, architectures, and metrics. The proposed benchmark utilizes diverse public seismic datasets (F3 Netherlands, Penobscot, Parihaka) and standard performance metrics to provide a unified, objective, and reproducible evaluation framework for future seismic segmentation models.
 
 
References:
• Monteiro, B. A. A., Canguçu, G. L., Jorge, L. M. S., Vareto, R. H., Oliveira, B. S., Silva, T. H., Lima, L. A., Machado, A. M. C., Schwartz, W. R., & Vaz-de-Melo, P. O. S. (2024). Literature review on deep learning for the segmentation of seismic images. Earth-Science Reviews, 258, 104955. https://doi.org/10.1016/j.earscirev.2024.104955
• Canguçu, G. L., Jorge, L. M. S., Barreto, G. T., Silva, T. H., Lima, L. A., Caldas, W. S., Tavares, C. G. S., Schwartz, W. R., Vaz-de-Melo, P. O. S., & Machado, A. M. C. (n.d.). A Principled Benchmark for Seismic Data Segmentation. 
2024

License Plate Recognition

This research addresses the critical challenge of License Plate Recognition (LPR) in real-world scenarios, where low-resolution (LR) and low-quality images from surveillance systems often lead to characters blending with the background or neighboring characters, severely hindering accurate recognition. Existing super-resolution methods often degrade network accuracy by failing to differentiate similar characters (e.g., 'Q' and 'O', 'B' and '8') or focus solely on objective metrics like PSNR and SSIM, which do not correlate well with actual LPR performance. Furthermore, a significant problem is the scarcity of publicly available datasets with paired real-world LR and high-resolution (HR) LP images and standardized evaluation protocols, leading to reliance on private or synthetically degraded data that fails to replicate real-world noise patterns.
 
Motivated by the urgent need for robust, efficient, and accurate LPR systems for applications like traffic monitoring and forensic investigations, this work proposes novel super-resolution approaches. These include new perceptual loss functions (e.g., Layout and Character Oriented Focal Loss - LCOFL) that integrate character recognition and LP layout awareness, enhanced architectures utilizing PixelShuffle layers, deformable convolutions, and attention modules for improved feature extraction and reorganization, and the introduction of UFPR-SR-Plates, a new public dataset and benchmark using real-world data and multi-image fusion strategies to provide a more realistic and rigorous evaluation framework.
 
 
References:
• Nascimento, V., Laroca, R., Lambert, J. d. A., Schwartz, W. R., & Menotti, D. (2022). Combining attention module and pixel shuffle for license plate super-resolution. Conference on Graphics, Patterns and Images (SIBGRAPI), 228–233. https://doi.org/10.1109/SIBGRAPI55357.2022.9991753
• Nascimento, V., Laroca, R., Ribeiro, R. O., Schwartz, W. R., & Menotti, D. (2024). Enhancing license plate super-resolution: A layout-aware and character-driven approach. Conference on Graphics, Patterns and Images (SIBGRAPI) 2024. https://doi.org/10.1109/SIBGRAPI62404.2024.10716303
• Nascimento, V., Lima, G. E., Ribeiro, R. O., Schwartz, W. R., Laroca, R., & Menotti, D. (2025). Toward Advancing License Plate Super-Resolution in Real-World Scenarios: A Dataset and Benchmark. Journal of the Brazilian Computer Society. https://valfride.github.io/nascimento2024toward/
• Nascimento, V., Laroca, R., Lambert, J. d. A., Schwartz, W. R., & Menotti, D. (2023). Super-Resolution of License Plate Images Using Attention Modules and Sub-Pixel Convolution Layers. Computers & Graphics, 113, 69–76. https://doi.org/10.1016/j.cag.2023.05.005
• Gonçalves, G. R., Diniz, M. A., Laroca, R., Menotti, D., & Schwartz, W. R. (2019). Multi-Task Learning for Low-Resolution License Plate Recognition. In Iberoamerican Congress on Pattern Recognition (CIARP), 251–261. https://doi.org/10.1007/978-3-030-33904-3_23
2021

Deep Learning Model Pruning

This research focuses on Deep Network Compression for Convolutional Neural Networks (CNNs), tackling the critical problem of their high computational cost, energy consumption, and memory requirements that hinder deployment on resource-constrained systems like mobile and IoT devices. A key motivation is to overcome limitations of existing pruning methods that often degrade network accuracy by indiscriminately removing filters or demand extensive effort.

The work proposes novel discriminative structured pruning strategies, including filter pruning and layer pruning, both leveraging Partial Least Squares (PLS) to identify and remove less important components while preserving or improving network accuracy. This aims to achieve superior resource-efficiency, including reduced FLOPs, memory usage, and prediction time, by intelligently reducing network width and depth.

References:
• Jordao, A., Yamada, F., & Schwartz, W. R. (2020). Deep Network Compression based on Partial Least Squares. Neurocomputing.
• Jordao, A., Lie, M., & Schwartz, W. R. Discriminative Layer Pruning for Convolutional Neural Networks