PPPascal Penava

Pascal Penava

*This is an algorithmically generated author record based on publicly available information.

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


Pascal Penava has a total of 20 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    A Novel High-Performance Deep-Learning-Based Approach for the Detection of Arsenic Poisoning

    by Norbert Moenks, Pascal Penava, Ricardo Buettner

    Arsenic poisoning remains a serious public health problem that affects millions of people worldwide. Early detection is crucial to prevent severe long-term complications and implement mitigation strategies as recommended by the World Health Organization. While traditional assessment methods require specialists and laboratory equipment, deep learning methods have shown potential for fast and automated analysis of skin images to detect its dermatological manifestations, also called arsenicosis. However, existing approaches suffer from insufficient validation and suboptimal detection performance. To address this problem, we present a novel deep learning-based approach to classify skin images into two categories: healthy and arsenic affected, using a compact EfficientNet backbone with a novel pre-processing step that suppresses irrelevant high-frequency artifacts while retaining lesion morphology and lesion-to-skin transitions characteristic of arsenicosis. With a balanced accuracy of 99.93%, our work sets a new benchmark for the binary classification of arsenicosis, reducing the error rate by 79.5% compared to the previous benchmark. These results indicate that combining specific image enhancements with deep learning can improve the analysis of arsenic-affected skin images. The pipeline is lightweight, easily reproducible, and transferable to real-world scenarios such as mobile or telemedical applications, enabling reliable, non-invasive screening and monitoring in arsenic-affected populations.


  • Open Access

    A Novel Hybrid Transformer Approach for Text-Based Defect Classifications in the Construction Industry

    by Ricardo Buettner, Pascal Penava, Max Wiedemann

    In the construction industry, defects and their rework cause significant unplanned overheads. Defects are recorded in natural language, and while traditional manual evaluation and classification of defects is labor- and resource-intensive, transformer-based deep learning methods enable automated and efficient classification. Previous automated defect classification systems have the problem that their classification accuracy decreases with broad coverage of diverse defect types. To address the complexity and diversity of different defects and classify them correctly, we use a novel hybrid transformer-based deep learning approach (all-MiniLM-L12-v2 and all-mpnet-base-v2) to enable the classification of defect descriptions in construction projects. The hybrid approach enables advanced analysis of texts with different focuses, allowing for simultaneous addressing of a broad coverage of defect types.With an average accuracy of 90.04% across all classes, our model sets the new benchmark for defect classification in construction projects, using the example of fire door inspections. The results show that the novel hybrid approach with two transformer architectures can significantly improve the classification of defect descriptions. However, the results could be influenced by the subjectivity of the labels generated by experts, and by an "others" class whose broadness might limit the precision of the classification. The approach shown can be easily embedded in real construction projects and thus serves as an essential first step toward automated defect handling.


  • Open Access

    A High-Performance Deep-Learning-Based Ground Penetrating Radar Classification Approach With Special Focus on Multi-Class Robustness

    by Mats Pagers, Pascal Penava, Ricardo Buettner

    Urbanization is increasing the presence of various utilities in the ground, while at the same time underground cavities and sinkholes pose a major threat to the civilization and structural integrity of the infrastructure. Ground Penetrating Radar B-scans enable the non-invasive and mobile analysis of the ground. While manual analysis methods are complex and labor-intensive, deep learning methods have shown potential for automated analysis of the scans. While existing models allow a binary classification, this study uses a novel deep learning-based approach to classify Ground Penetrating Radar B-scan radargrams into three subsurface categories (utilities, cavities, and intact zones) to account for the complexity of real-world subsurface conditions. A ResNet50 model is used with a novel edge preserving preprocessing step to enhance the quality of input data, with the aim to reduce noise in the radargrams while keeping important structural details unblurred. The model shows significant improvements when domain-specific filtering is applied. Balanced accuracy increased from 93.06 % to 96.00 %, setting a new benchmark for Ground Penetrating Radar-based multi-class classification. The results show that combining image enhancement with deep learning can improve the analysis of underground imagery. The method is easily reproducible and is applicable in real-world scenarios such as infrastructure maintenance, utility mapping or archaeological prospection.

  • Open Access

    Deep-Learning-Based Determination of Textile Properties: A Novel Triplet Architecture Approach for Classifying Cotton Content

    by Max Wiedemann, Pascal Penava, Christopher Mai, Ricardo Buettner

    To reach sustainable production and consumption patterns, recycling is a key task. Automating the recycling process, especially in sorting tasks, is a strong hope to improve the efficiency and economic viability of the recycling industry. Especially the determination of textile properties such as cotton content is an important task in recycling and the textile industry as a whole. Approaches using NIR-spectrography are common, but can be costly and produce complex datasets. We therefore propose a visual approach to classify fabrics after their cotton content. For this, we apply a novel triplet-architecture approach with a variety of modifications. We combine ResNet50, EfficientNetB4, and DenseNet121 in this architecture in order to make use of their respective strengths while overcoming the weaknesses of the models. For a further accuracy enhancement, we modified all three architectures with adaptive feature pyramid networks, and we added a deformable convolution layer to ResNet50 and DenseNet121. Additionally, we use a second fully connected layer to enhance the model’s classification capacity. With this architecture approach, we achieve an average Root Mean Squared Error of 14.77%, setting a new benchmark for visual approaches to cotton-content classification using cross-validation. We further prove the effectiveness and enhanced accuracy of using a triplet model approach, as well as using all of our modifications. Visual approaches are not market-ready yet, but we show the potential of deep-learning methods for lowering labor-intensiveness and time needed in textile property determination. By that, the recycling industry can be made more efficient and economically viable, helping to come closer to circular economy practices in the textile industry.


  • Open Access

    A Systematic Literature Review of Large Language Model Applications in Industry

    by Norbert Moenks, Pascal Penava, Ricardo Buettner

    Large Language Models are rapidly transforming processes across industries by enabling advanced capabilities in natural language understanding, code generation, diagnostics, and decision support. Despite the growing adoption of this technology, a systematic understanding of their application across the industrial value creation processes remains lacking. This paper addresses this gap by conducting a systematic literature review of 96 peer-reviewed studies, following the PRISMA guidelines. Based on this foundation, large language model use cases across industries were identified, categorized, and structured using the primary and secondary activities of Porter’s value chain as a classification framework. The analysis reveals that LLM adoption is heavily concentrated in technology-focused and internal operational activities across industries, where they offer immediate benefits at lower risk. In contrast, areas such as logistics, procurement, and customer-facing functions remain largely unexplored, mainly due to challenges related to integration, data governance, and regulatory requirements. The analysis shows that current deployments are primarily limited to isolated, manageable use cases, leaving substantial innovation potential unrealized in underexplored value chain activities. These findings provide a foundation for further research and for the strategic adoption of large language models throughout the industrial value chain.


  • Open Access

    Brain Inspired Deep Learning Architecture For Non-Severe Depression Detection

    by Pascal Penava

    Based on the work of Penava & Buettner [1], in which they showed that their deep learning method sets a new benchmark for resting-state electroencephalogram based detection of nonsevere depression, I extend this work by transferring the communication-through-coherence idea of the human brain to a novel deep learning approach and thus raised the benchmark set in this work for the detection of non-severe depression. The results of my work confirm and extend the finding that non-severe depressions can be detected with modern deep learning methods based on resting-state electroencephalogram data, and that there is a correlation between them.

  • Open Access

    A Systematic Literature Review of Machine Learning-Based Personality Trait Detection Using Electroencephalographic Data

    by Celina Rieck, Pascal Penava, Ricardo Buettner

    Traditional self-report personality assessments, such as detecting Big Five personality traits through the NEO-Five-Factor-Inventory, are personnel-, time- and cost-intensive approaches that are prone to bias, whereas electroencephalography provides a biometric alternative by capturing neural correlates of personality. That is why machine learning-based detection of personality traits using electroencephalographic data has gained increasing attention. The integration of machine and deep learning algorithms has improved automated trait classification but the comparative reliability of these approaches remains unclear. This systematic literature review examines if trait detection is possible by using electroencephalography in combination with machine and deep learning models, analyzing 58 studies since 2015. We compare their performance by a meta-analysis of weighted data, highlighting strengths and limitations. Our findings indicate that machine and deep learning models achieve varying degrees of accuracy in detecting different personality traits, with Openness demonstrating the highest classification accuracy, while Neuroticism remains the most challenging to detect. Deep learning models, particularly hybrid architectures, outperform traditional machine learning classifiers, highlighting the advantage of deep feature extraction in EEG data processing. However, methodological inconsistencies, limited dataset standardization and the influence of transient emotional states pose significant challenges. Despite these limitations, machine and deep learning models based on electroencephalography show promise for applications in mental health, recruitment and personalized interventions. Further research is needed to standardize methods and improve model reliability for real-world application.


  • Open Access

    Advancements in Landmine Detection: Deep Learning-Based Analysis With Thermal Drones

    by Daniel Heuschmid, Oliver Wacker, Yannick Zimmermann, Pascal Penava, Ricardo Buettner

    The pervasive threat of landmines across conflict-affected regions necessitates advancements in detection technologies to enhance safety and efficiency in demining efforts. Furthermore, the development of a solution that can be effectively utilized in both resource-rich and resource-poor environments is essential, as many conflict-affected regions lack the financial means or technical infrastructure to deploy expensive and complex detection technologies. This paper introduces a deep learning-based approach utilizing uncrewed aerial vehicles equipped with thermal imaging cameras for landmine detection. Leveraging the MobileNetV3-Large architecture and building upon it, we propose a lightweight, yet powerful, machine learning model that can be seen as a safe, cost-efficient and fast landmine detection method. We evaluated our approach using a unique dataset comprising 2,700 thermographic images captured by a DJI Matrice 100 drone with a Zenmuse XT infrared camera. Through meticulous data pre-processing and augmentation strategies, we enhance the model’s ability to generalize across different terrains and mine types, addressing one of the primary challenges in landmine detection. The evaluation of our model on a rigorously curated test set demonstrates promising results, achieving a training accuracy of 96.97 %, a validation accuracy of 97.19 %, and a test accuracy of 96.14 %. These metrics not only demonstrate the model’s effectiveness in identifying landmines from thermal images but also highlight its potential to outperform other thermal approaches in real-world demining operations. This supports the application of uncrewed aerial vehicles and deep learning technologies in humanitarian and environmental challenges.


  • Open Access

    A Novel High Performance Object Identification Approach in Care Homes Using Gaussian Preprocessing

    by Sara Auweiler, Melina Mueller, Diana Puhla, Pascal Penava, Ricardo Buettner

    As the global population ages, the prevalence of dementia is expected to surge, necessitating innovative solutions to support the growing number of affected individuals. This research shows a innovative approach for the application of Convolutional Neural Networks for object identification in nursing home environments to enhance the independence and quality of life of dementia patients. Therefore we compare the performance of four well-known Convolutional Neural Network architectures — VGG16, VGG19, InceptionV3, and ResNet50 — to identify the most effective model for this application. Building on this, we use Gaussian blurring as an innovative preprocessing technique to improve model accuracy by reducing high-frequency noise and enhancing feature extraction. Our results indicate that ResNet50 outperforms other models, achieving the highest accuracy with 90.53%. Optimization through data augmentation and fine-tuning further enhances ResNet50’s performance to 94.69%. Applying the Gaussian filter significantly improves the model’s accuracy to 96.81%, demonstrating its potential as a crucial preprocessing step and setting a new benchmark for this application. Our findings provide valuable insight into applying the Gaussian filter on Convolutional Neural Networks and pave the way for developing intelligent support systems to assist dementia patients in their daily tasks and improve their overall well-being.