Ricardo Buettner
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
Research Subjects
Ricardo Buettner has a total of 99 co-authors. Here are the top co-authors by number of shared publications:
- Pascal Penava
Chair of Hybrid Intelligence, Helmut-Schmidt-University/University of the Federal Armed Forces Hamburg, Hamburg, Germany
17 shared publications - Luca Eisentraut
Affiliation Chair of Hybrid Intelligence, Helmut-Schmidt-University/University of the Federal Armed Forces Hamburg, Hamburg, Germany
16 shared publications - Christopher Mai
Chair of Hybrid Intelligence, Helmut-Schmidt-University/University of the Federal Armed Forces Hamburg, Hamburg, Germany
10 shared publications - Leopold Fischer-Brandies
Chair of Hybrid Intelligence, Helmut-Schmidt-University/University of the Federal Armed Forces Hamburg, Hamburg, Germany
7 shared publications Christopher Mai
Chair of Hybrid Intelligence, Helmut-Schmidt-University/University of the Federal Armed Forces Hamburg, Hamburg, Germany
5 shared publications- Max Wiedemann
Helmut-Schmidt-University/University of the Federal Armed Forces Hamburg
5 shared publications
Latest Publications
- Open Access
- 3 September 2026
- Published by Elsevier in Data in Brief
- Vol. 69
A multi-label sentinel-2 dataset for deep learning-based energy, transport, and storage infrastructure mapping in Germany
- 3 September 2026
- Published by Elsevier in Data in Brief
- Vol. 69
Open AccessA multi-label sentinel-2 dataset for deep learning-based energy, transport, and storage infrastructure mapping in Germany
- Open Access
- 28 August 2026
- Published by Elsevier in Data in Brief
- Vol. 68, 113215
DefectAtlas-20: A real-world image dataset for joint material and surface-condition classification
- 28 August 2026
- Published by Elsevier in Data in Brief
- Vol. 68, 113215
Open AccessDefectAtlas-20: A real-world image dataset for joint material and surface-condition classification
- Open Access
- 24 July 2026
- Published by Elsevier in Data in Brief
- Vol. 68
A Sentinel-2 dataset for railway presence classification across metropolitan France
- 24 July 2026
- Published by Elsevier in Data in Brief
- Vol. 68
Open AccessA Sentinel-2 dataset for railway presence classification across metropolitan France
- Open Access
- 16 July 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (7) , e0353374
A novel deep-learning approach for robust identification of plant diseases.
Rising temperatures and changing weather conditions are accelerating the spread of plant diseases and increasing the threat to global food security. Reliable detection of leaf diseases is therefore essential to protect crop yields and ensure food quality. Deep learning has proven to be a powerful tool for classifying leaf diseases across various crops. Due to the natural variability of plants, plant diseases often appear in irregular structures. Surface unevenness, folds, or dirt particles are common in field images and can be mistakenly identified as important features by convolutional neural networks (CNNs). This is a challenge that has not been sufficiently addressed in previous studies. This study proposes a novel deep learning approach that takes into account both the specific visual characteristics of plant diseases and potential disturbances in the microstructure, such as surface irregularities or prominent leaf veins, which may mislead the model. Using stratified five-fold cross-validation on a peer-reviewed dataset, which comprises 2,801 images of radish leaves across five classes (healthy, three disease classes: mosaic virus, black leaf spot, and downy mildew, and one pest-affected class: flea beetle), the proposed method achieved an average and balanced accuracy of 99.86%, establishing a new dataset-level benchmark in the field and demonstrating its effectiveness. The results indicate that the proposed approach may provide a promising basis for future applications in agricultural field monitoring, automated sorting and post-harvest quality control, offering potential to reduce both food waste and associated costs.
- 16 July 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (7) , e0353374
Open AccessA novel deep-learning approach for robust identification of plant diseases.
Rising temperatures and changing weather conditions are accelerating the spread of plant diseases and increasing the threat to global food security. Reliable detection of leaf diseases is therefore essential to protect crop yields and ensure food quality. Deep learning has proven to be a powerful tool for classifying leaf diseases across various crops. Due to the natural variability of plants, plant diseases often appear in irregular structures. Surface unevenness, folds, or dirt particles are common in field images and can be mistakenly identified as important features by convolutional neural networks (CNNs). This is a challenge that has not been sufficiently addressed in previous studies. This study proposes a novel deep learning approach that takes into account both the specific visual characteristics of plant diseases and potential disturbances in the microstructure, such as surface irregularities or prominent leaf veins, which may mislead the model. Using stratified five-fold cross-validation on a peer-reviewed dataset, which comprises 2,801 images of radish leaves across five classes (healthy, three disease classes: mosaic virus, black leaf spot, and downy mildew, and one pest-affected class: flea beetle), the proposed method achieved an average and balanced accuracy of 99.86%, establishing a new dataset-level benchmark in the field and demonstrating its effectiveness. The results indicate that the proposed approach may provide a promising basis for future applications in agricultural field monitoring, automated sorting and post-harvest quality control, offering potential to reduce both food waste and associated costs.
- Open Access
- 4 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 86797-86818
A Systematic Literature Review of Prompt Injection Attacks in LLM-Integrated Systems
Prompt injection exploits a fundamental property of large language models: the inability to distinguish trusted instructions from untrusted content when both are processed as natural language tokens. As these models are integrated into tool-augmented and memory-enabled systems, the attack surface expands beyond the prompt interface, yet no existing review maps this landscape against standardised adversarial frameworks or analyses the balance between offensive and defensive research. This review follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. We searched IEEE Xplore, ACM Digital Library, and ScienceDirect for peer-reviewed studies published between 2015 and October 2025, yielding 89 included publications of which 81 form the analytical corpus. Aligned to the MITRE Adversarial Threat Landscape for AI Systems framework (version 5.4.0), each paper was coded along three axes: trust boundary, attacker objective, and research orientation with attacker objectives further classified into seventeen operational patterns aligned to MITRE ATLAS techniques. The corpus is dominated by research on the prompt interface and by two attacker objectives, Instruction Hijack and Policy Bypass, leaving underexplored boundaries lacking both attack characterisation and defence proposals. This architectural narrowness does not reflect the multi-component architecture of production LLM systems. The three-axis framework and coded corpus establish a baseline for tracking how coverage evolves as deployment architectures expand. The concentration reflects a research incentive structure that favours reproducible prompt-level experiments over architecturally complex assessments at tool and memory boundaries. The field exhibits a maturity gradient in which defence development is concentrated at the prompt interface while other boundaries remain in early or absent characterisation. The corpus treats prompt injection as a model-level property rather than a system-level vulnerability.
- 4 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 86797-86818
Open AccessA Systematic Literature Review of Prompt Injection Attacks in LLM-Integrated Systems
Prompt injection exploits a fundamental property of large language models: the inability to distinguish trusted instructions from untrusted content when both are processed as natural language tokens. As these models are integrated into tool-augmented and memory-enabled systems, the attack surface expands beyond the prompt interface, yet no existing review maps this landscape against standardised adversarial frameworks or analyses the balance between offensive and defensive research. This review follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. We searched IEEE Xplore, ACM Digital Library, and ScienceDirect for peer-reviewed studies published between 2015 and October 2025, yielding 89 included publications of which 81 form the analytical corpus. Aligned to the MITRE Adversarial Threat Landscape for AI Systems framework (version 5.4.0), each paper was coded along three axes: trust boundary, attacker objective, and research orientation with attacker objectives further classified into seventeen operational patterns aligned to MITRE ATLAS techniques. The corpus is dominated by research on the prompt interface and by two attacker objectives, Instruction Hijack and Policy Bypass, leaving underexplored boundaries lacking both attack characterisation and defence proposals. This architectural narrowness does not reflect the multi-component architecture of production LLM systems. The three-axis framework and coded corpus establish a baseline for tracking how coverage evolves as deployment architectures expand. The concentration reflects a research incentive structure that favours reproducible prompt-level experiments over architecturally complex assessments at tool and memory boundaries. The field exhibits a maturity gradient in which defence development is concentrated at the prompt interface while other boundaries remain in early or absent characterisation. The corpus treats prompt injection as a model-level property rather than a system-level vulnerability.
- Open Access
- 29 May 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (5) , e0349562
Towards intelligent railway monitoring: A novel hybrid deep learning architecture for railway obstacle detection
Railways are among the most efficient modes of transportation, capable of moving large quantities of goods and passengers over long distances at relatively low cost. However, accidents frequently occur due to objects or individuals present on the tracks, as trains are unable to swerve and require long braking distances. While the localization of objects within the track bed is a well-explored topic, the reliable and high-performance classification of such obstacles across all relevant categories remains an unresolved challenge. This study proposes an innovative hybrid architecture that leverages the specific visual characteristics of track bed imagery, setting a new benchmark in this domain. The hybrid design effectively leverages the strengths of ResNet50 and Swin Transformer V2, allowing the model to capture both local and global features. In addition, an Efficient Attention Module is integrated to further emphasize the most relevant features for robust obstacle classification. Using stratified five-fold cross-validation on a dataset of 2,003 images across six classes (iron bar, boulder, person, branch, canister, and barrel), the model achieved an average balanced accuracy of 99.46%. The results have implications for accident prevention, improving operational efficiency, and modernizing railway safety systems, thereby enabling the future application of automatic railway surveillance systems to ultimately enhance operational security.
- 29 May 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (5) , e0349562
Open AccessTowards intelligent railway monitoring: A novel hybrid deep learning architecture for railway obstacle detection
Railways are among the most efficient modes of transportation, capable of moving large quantities of goods and passengers over long distances at relatively low cost. However, accidents frequently occur due to objects or individuals present on the tracks, as trains are unable to swerve and require long braking distances. While the localization of objects within the track bed is a well-explored topic, the reliable and high-performance classification of such obstacles across all relevant categories remains an unresolved challenge. This study proposes an innovative hybrid architecture that leverages the specific visual characteristics of track bed imagery, setting a new benchmark in this domain. The hybrid design effectively leverages the strengths of ResNet50 and Swin Transformer V2, allowing the model to capture both local and global features. In addition, an Efficient Attention Module is integrated to further emphasize the most relevant features for robust obstacle classification. Using stratified five-fold cross-validation on a dataset of 2,003 images across six classes (iron bar, boulder, person, branch, canister, and barrel), the model achieved an average balanced accuracy of 99.46%. The results have implications for accident prevention, improving operational efficiency, and modernizing railway safety systems, thereby enabling the future application of automatic railway surveillance systems to ultimately enhance operational security.
- Open Access
- 5 May 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (5) , e0346583
A hybrid deep-learning-architecture for identifying cotton content in fabric materials
Recycling plays a crucial role in achieving sustainable production. In particular, automating sorting processes holds great promise for enhancing both the efficiency and economic feasibility of the recycling industry. One challenge within this context is the classification of fabrics based on their cotton content. This task is relevant not only for recycling but also for the broader textile sector. Traditional methods often rely on manual labor, which is both time-consuming and labor-intensive, while advanced techniques like near-infrared spectrography, although effective, can be complex and expensive. We therefore propose a task-specific, deep-learning-based hybrid architecture approach for visually classifying fabrics based on their cotton content. The hybrid architecture leverages the strengths of DenseNet121 and Swin Transformer V2. The hybrid network is capable of capturing both local and global features, which enables it to detect differences in fiber types as well as quantify their presence within the fabric. To enhance its classification accuracy, we modified DenseNet121 with an adaptive feature pyramid network, which helps to consider features extracted at different levels, and a deformable convolution layer, focusing on structures in the fabric. Stratified 5-fold cross-validation was employed on a peer-reviewed dataset to assess the model’s performance and ensure its robustness. Compared to the state-of-the-art, we set a new benchmark for cross-validated visual approaches using standard camera imagery for cotton content classification with an average Root Mean Squared Error of 14.01%. We therefore prove the effectiveness of our architecture and its modifications. Our approach demonstrates the potential benefits of using deep learning methods for determining cotton content. These methods can help reduce manual effort, lower costs, and ultimately improve the economic situation of recycling companies.
- 5 May 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (5) , e0346583
Open AccessA hybrid deep-learning-architecture for identifying cotton content in fabric materials
Recycling plays a crucial role in achieving sustainable production. In particular, automating sorting processes holds great promise for enhancing both the efficiency and economic feasibility of the recycling industry. One challenge within this context is the classification of fabrics based on their cotton content. This task is relevant not only for recycling but also for the broader textile sector. Traditional methods often rely on manual labor, which is both time-consuming and labor-intensive, while advanced techniques like near-infrared spectrography, although effective, can be complex and expensive. We therefore propose a task-specific, deep-learning-based hybrid architecture approach for visually classifying fabrics based on their cotton content. The hybrid architecture leverages the strengths of DenseNet121 and Swin Transformer V2. The hybrid network is capable of capturing both local and global features, which enables it to detect differences in fiber types as well as quantify their presence within the fabric. To enhance its classification accuracy, we modified DenseNet121 with an adaptive feature pyramid network, which helps to consider features extracted at different levels, and a deformable convolution layer, focusing on structures in the fabric. Stratified 5-fold cross-validation was employed on a peer-reviewed dataset to assess the model’s performance and ensure its robustness. Compared to the state-of-the-art, we set a new benchmark for cross-validated visual approaches using standard camera imagery for cotton content classification with an average Root Mean Squared Error of 14.01%. We therefore prove the effectiveness of our architecture and its modifications. Our approach demonstrates the potential benefits of using deep learning methods for determining cotton content. These methods can help reduce manual effort, lower costs, and ultimately improve the economic situation of recycling companies.
- Open Access
- 29 April 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (4) , e0346708
DualFusionNet: A fusion-based dual architecture for visual quality control on fabric surfaces
Zippers are essential in fashion and beyond. Found on jackets, shoes, and bags, they provide a versatile and practical opening and closing mechanism that seamlessly integrates into everyday use and specialized applications. The quality control of zippers is often still based on a subjective visual inspection by experienced persons. This step is time-consuming, inefficient, and associated with a higher error rate. To address these challenges, the deployment of image processing systems for the automated inspection of zippers is becoming increasingly crucial. To determine whether defects are present, approaches often use architectures that perform local feature extraction. However, there is a risk that global relationships will be overlooked, which must not be ignored in the case of defects. The aim of this study is therefore to improve the detection accuracy of zipper defects by integrating both local and global characteristics into one architecture. To achieve this aim, DualFusionNet is proposed, a dual architecture that combines the advantages of CNN- and Transformer-based feature extraction. The integration of the Adaptive Feature Pyramid Network (AFPN) with a Squeeze-and-Excitation (SE) block facilitates the fusion of features from both architectures, enabling the dual architecture to incorporate local and global relationships for improved classification performance. The efficacy of this approach is evidenced by the attainment of an accuracy of 99.74% and a balanced accuracy of 99.58%. The high accuracy shows that the use of deep learning systems for the automatic inspection of defects in zippers is an effective economic step for companies, as it leads to cost and time savings in production due to the high classification performance. This study provides an overview of the results achieved and other key performance indicators of the architecture used.
- 29 April 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (4) , e0346708
Open AccessDualFusionNet: A fusion-based dual architecture for visual quality control on fabric surfaces
Zippers are essential in fashion and beyond. Found on jackets, shoes, and bags, they provide a versatile and practical opening and closing mechanism that seamlessly integrates into everyday use and specialized applications. The quality control of zippers is often still based on a subjective visual inspection by experienced persons. This step is time-consuming, inefficient, and associated with a higher error rate. To address these challenges, the deployment of image processing systems for the automated inspection of zippers is becoming increasingly crucial. To determine whether defects are present, approaches often use architectures that perform local feature extraction. However, there is a risk that global relationships will be overlooked, which must not be ignored in the case of defects. The aim of this study is therefore to improve the detection accuracy of zipper defects by integrating both local and global characteristics into one architecture. To achieve this aim, DualFusionNet is proposed, a dual architecture that combines the advantages of CNN- and Transformer-based feature extraction. The integration of the Adaptive Feature Pyramid Network (AFPN) with a Squeeze-and-Excitation (SE) block facilitates the fusion of features from both architectures, enabling the dual architecture to incorporate local and global relationships for improved classification performance. The efficacy of this approach is evidenced by the attainment of an accuracy of 99.74% and a balanced accuracy of 99.58%. The high accuracy shows that the use of deep learning systems for the automatic inspection of defects in zippers is an effective economic step for companies, as it leads to cost and time savings in production due to the high classification performance. This study provides an overview of the results achieved and other key performance indicators of the architecture used.
- Open Access
- 3 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 56138-56154
RobustRailNet: Architecture-Driven Adversarial Robustness for Safety-Critical Railway Obstacle Perception
State-of-the-art architectures for automated railway safety surveillance achieve high benchmark accuracies; however, such systems are systematically vulnerable to adversarial perturbations. Current approaches focus on adversarial training, while other aspects, such as architecture design, are overlooked as robustness mechanisms. We propose RobustRailNet, a novel architecture for the robust and high-performance classification of railway obstacles. The proposed approach builds upon domain-relevant characteristics such as high-frequency textures, spatially coherent structures, and global context. RobustRailNet is evaluated against seven white-box adversarial attacks, including gradient-based, optimization-based, and patch-based methods, representing a worst-case assumption for model vulnerability. The evaluation is conducted on a diverse dataset comprising multiple railway obstacle types. RobustRailNet achieves a highly significant mean robustness gain of up to 66.89 percentage points compared to the state-of-the-art while maintaining a mean balanced accuracy of 99.52%. The results demonstrate that adversarial vulnerability arises from architectural design and is not exclusively linked to training data or strategies. Architectural optimization should therefore be seen as a necessity rather than a possibility for the design of robust, safety-critical systems. When developing and deploying such architectures in practice, the influence of the architecture should be taken into account upfront, rather than attempting to achieve robustness exclusively through ex post retraining.
- 3 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 56138-56154
Open AccessRobustRailNet: Architecture-Driven Adversarial Robustness for Safety-Critical Railway Obstacle Perception
State-of-the-art architectures for automated railway safety surveillance achieve high benchmark accuracies; however, such systems are systematically vulnerable to adversarial perturbations. Current approaches focus on adversarial training, while other aspects, such as architecture design, are overlooked as robustness mechanisms. We propose RobustRailNet, a novel architecture for the robust and high-performance classification of railway obstacles. The proposed approach builds upon domain-relevant characteristics such as high-frequency textures, spatially coherent structures, and global context. RobustRailNet is evaluated against seven white-box adversarial attacks, including gradient-based, optimization-based, and patch-based methods, representing a worst-case assumption for model vulnerability. The evaluation is conducted on a diverse dataset comprising multiple railway obstacle types. RobustRailNet achieves a highly significant mean robustness gain of up to 66.89 percentage points compared to the state-of-the-art while maintaining a mean balanced accuracy of 99.52%. The results demonstrate that adversarial vulnerability arises from architectural design and is not exclusively linked to training data or strategies. Architectural optimization should therefore be seen as a necessity rather than a possibility for the design of robust, safety-critical systems. When developing and deploying such architectures in practice, the influence of the architecture should be taken into account upfront, rather than attempting to achieve robustness exclusively through ex post retraining.
- Open Access
- 23 January 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (1) , e0340789
DualSightNet: A novel dual architecture for visual quality control of railway infrastructure
To ensure the operational safety of trains, it is essential to monitor the condition of the rails. In order to detect problems ranging from wear and tear to possible sabotage, no comprehensive and continuous monitoring is carried out today. The objective of this study is to develop and robustly validate a deep-learning model capable of reliably classifying a broad range of safety-critical rail surface defects. The use of deep learning to classify defect types based on optical images is a promising approach, but the existing literature does not yet achieve a robust, high-performing classification for a broad range of failure types. Many approaches rely mainly on local feature extraction, but this carries the risk of overlooking global relationships, which are crucial for distinguishing certain defect types. This study addresses this gap by utilizing the specific features of this problem domain (small, local, and global defect types). To this end, DualSightNet is introduced as a hybrid architecture enhanced by an attention module for classifying a broad range of railway track surface defects. The model achieves a five-fold cross-validated average balanced accuracy of 97.55 % on a peer-reviewed, real-world dataset of 5,153 images covering seven defect types, recorded directly from an inspection vehicle under operational conditions, indicating strong generalization across the diverse real-world variations represented in the dataset. Compared to existing CNN- or Transformer-based approaches, DualSightNet is the first approach that combines local and global feature extraction through a gating-based fusion mechanism and then enhances the fused representation using an attention module, which enables substantially more robust multiclass defect recognition. This sets a new benchmark for our problem domain, surpassing previous approaches, which either lack broad defect coverage or do not employ rigorous cross-validated evaluation. Our results have far-reaching practical implications, proving that by leveraging problem-specific features, neural networks are able to robustly classify a broad range of defect types. The inference time of the proposed system (3.00 ms per image) makes DualSightNet suitable for deployment in automated inspection vehicles and real-time monitoring scenarios.
- 23 January 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (1) , e0340789
Open AccessDualSightNet: A novel dual architecture for visual quality control of railway infrastructure
To ensure the operational safety of trains, it is essential to monitor the condition of the rails. In order to detect problems ranging from wear and tear to possible sabotage, no comprehensive and continuous monitoring is carried out today. The objective of this study is to develop and robustly validate a deep-learning model capable of reliably classifying a broad range of safety-critical rail surface defects. The use of deep learning to classify defect types based on optical images is a promising approach, but the existing literature does not yet achieve a robust, high-performing classification for a broad range of failure types. Many approaches rely mainly on local feature extraction, but this carries the risk of overlooking global relationships, which are crucial for distinguishing certain defect types. This study addresses this gap by utilizing the specific features of this problem domain (small, local, and global defect types). To this end, DualSightNet is introduced as a hybrid architecture enhanced by an attention module for classifying a broad range of railway track surface defects. The model achieves a five-fold cross-validated average balanced accuracy of 97.55 % on a peer-reviewed, real-world dataset of 5,153 images covering seven defect types, recorded directly from an inspection vehicle under operational conditions, indicating strong generalization across the diverse real-world variations represented in the dataset. Compared to existing CNN- or Transformer-based approaches, DualSightNet is the first approach that combines local and global feature extraction through a gating-based fusion mechanism and then enhances the fused representation using an attention module, which enables substantially more robust multiclass defect recognition. This sets a new benchmark for our problem domain, surpassing previous approaches, which either lack broad defect coverage or do not employ rigorous cross-validated evaluation. Our results have far-reaching practical implications, proving that by leveraging problem-specific features, neural networks are able to robustly classify a broad range of defect types. The inference time of the proposed system (3.00 ms per image) makes DualSightNet suitable for deployment in automated inspection vehicles and real-time monitoring scenarios.