Jorge Carlos Santos Cardoso

Quality control of natural cork stoppers is essential in the wine industry, as aesthetic grading dictates market value. Current pricing relies heavily on visual quality, defining market classes ranging from premium “flor” to low-grade “quintos”. While traditional manual inspection is subjective, the application of supervised learning is strictly hindered by the high natural heterogeneity of cork and the unreliability of subjective ground-truth labels. To address this, we shift from rigid classification to an exploratory data analysis approach, proposing an unsupervised representation learning framework based on a Vision Transformer Autoencoder (ViTAE). The model leverages transformer-based feature extraction to obtain rich, multiview embeddings of cork surfaces. Evaluated on a large-scale industrial dataset (80,000 training and 40,000 test samples), the ViTAE latent space was compared against handcrafted descriptors (LBP and HOG) and autoencoder baselines (CAE and VAE). Performance was assessed via internal clustering metrics, statistical significance tests, and computational feasibility for real-time inspection. Results demonstrate that while classical descriptors like HOG produce statistically compact clusters, ViTAE yields a robust feature space with a critical operational advantage: spatial interpretability. Through image reconstruction, ViTAE acts as a visual auditing tool, where the reconstruction behavior naturally highlights atypical structural variations. This establishes ViTAE not merely as an automated classifier, but as a highly descriptive feature extractor capable of mapping raw material variability to support human-in-the-loop quality assessment systems.

Visual classification of natural cork bottle stoppers is a challenging problem due to the high similarity among non-premium aesthetic classes, where boundaries are subjective. Imperceptible variations in texture frequently overlap across categories, making fully supervised solutions expensive and unreliable, particularly when labeled data are scarce or ambiguous. In this context, we propose a data-efficient clustering-based model designed to support exploratory analysis and assist the labeling process by reducing uncertainty through confidence-aware decisions. The proposed framework consists of three stages. First, features are extracted using a Vision Transformer–based autoencoder, capturing both local texture details and global structural information from augmented data images. These latent representations are then refined through contrastive learning, improving discrimination among visually similar samples while preserving semantic structure. Finally, a fuzzy clustering matrix is generated from the refined features, making it suitable for handling uncertainties in industrial quality inspection. Experiments were conducted on real image data acquired from an operational cork stopper manufacturing plant, considering both a fully unsupervised scenario and a complementary supervised scenario in which labels were available for clearly separable classes. Performance was assessed using standard unsupervised clustering indices and Fuzzy Partition Coefficient, as well as supervised metrics such as clustering accuracy, normalized mutual information, and adjusted Rand index. Statistical analyses confirm that the proposed model achieves robust and competitive results when compared with state-of-the-art self-supervised learning approaches, demonstrating its practical relevance for industrial decision support under high uncertainty.

High-performance musical instrument training is a demanding discipline that engages cognitive, neurological, and physical skills. Professional musicians invest substantial time and effort into mastering their repertoire and developing the muscle memory and reflexes required to perform complex works in high-stakes settings. While existing surveys have explored the use of music in therapeutic and general training contexts, there is a notable lack of work focused specifically on the needs of professional musicians and advanced instrumental practice. This topical review explores the potential of EEG-based brain–computer interface (BCI) technologies to integrate real-time feedback of biomechanic and cognitive features in advanced musical practice. Building on a conceptual framework of technology-enhanced musical practice (TEMP), we review empirical studies of broad contexts, addressing the EEG signal decoding of biomechanic and cognitive tasks that closely relates to the specified TEMP features (movement and muscle activity, posture and balance, fine motor movements and dexterity, breathing control, head and facial movement, movement intention, tempo processing, ptich recognition, and cognitive engagement), assessing their feasibility and limitations. Our analysis highlights current gaps and provides a foundation for future development of BCI-supported musical training systems to support high-performance instrumental practice.

High-performance musical instrument training is a demanding discipline that engages cognitive, neurological, and physical skills. Professional musicians invest substantial time and effort into mastering repertoire and developing the muscle memory and reflexes required to perform complex works in high-stakes settings. While existing surveys have explored the use of music in therapeutic and general training contexts, there is a notable lack of work focused specifically on the needs of professional musicians and advanced instrumental practice. This topical review explores the potential of EEG-based brain-computer interface (BCI) technologies to integrate real-time feedback of biomechanic and cognitive features in advanced musical practice. Building on a conceptual framework of technology enhanced musical practice (TEMP), we review empirical studies of broad contexts, addressing EEG signal decoding of biomechanic and cognitive tasks that closely relate to the specified TEMP features (movement and muscle activity, posture and balance, fine motor and dexterity, breathing control, head and facial movement, movement intention, tempo processing, ptich recognition and cognitive engagement), assessing their feasibility and limitations. Our analysis highlights current gaps and provides a foundation for future development of BCI-supported musical training systems to support high-performance instrumental practice.

This work follows on from previous work on 3D photogrammetric modelling of rock masses, using the Structure from motion (SfM) technique and dense correlation as a basis. Photogrammetric 3D modelling is a technique that uses photographs to create three-dimensional models of objects or environments. In the context of rock masses, this approach is especially relevant for understanding the geometry and structure of these formations.Based on this concept, this work was developed with the integration of techniques, photogrammetric data acquisition and 3D photogrammetric modelling, from images acquired with an Unmanned Aerial Vehicle (UAV), and virtual reality (VR) technology, allowing users to explore the 3D model interactively. In this way, a realistic virtual environment is created, with the feeling of being present and experiencing a full-scale simulated environment, which makes it possible to visualise the rock mass from different perspectives and assess its structural geological characteristics and monitoring.The application of this technique to other areas of study (environmental changes and natural hazards), using other sensors (multispectral/hyperspectral optical, thermal sensors, LIDAR) that allow techniques for data fusion, is being studied considering the development of multi-platform and inter-disciplinary surveillance.

Science fiction movies have been a source of entertainment for many decades, but they have also been an inspiration for many people in the science and technology domains. Science fiction is not just about entertainment: it has been recently adopted also as a speculative design approach useful for designing and reflecting on the possibilities of future technologies. In this paper, we take existing science fiction works (movies) as speculative designs to understand and reflect on the possibilities of virtual reality. We have created a database of science fiction movies, TV shows, and episodes and analysed it along the dimensions of technology, form factor, user’s body position, and application domain. The result is a structured analysis of how existing science fiction movies have imagined the possibilities for virtual reality. This analysis may serve as further inspiration or for identifying research and development paths related to virtual reality.

We developed and assessed the Marine Pollution Control Simulator VR, a virtual reality training platform designed to improve the preparedness of marine pollution control teams. Traditional training for marine oil spills is costly and poses logistical and safety challenges. Our VR-based solution offers an immersive environment for training in oil spill containment and recovery, featuring navigation and equipment manipulation. The project emphasizes collaborative emergency response simulations, particularly vessel coordination. A user study evaluated the simulator’s usability, immersion, and potential for simulator sickness. Participants found the VR environment immersive and beneficial for training, though usability and physical discomfort issues were noted. These findings highlight the need for further refinements to enhance user experience and reduce adverse effects.