Robust Deep Learning for Real-World Visual Big Data

Computer vision and deep learning are increasingly applied to large-scale visual data across scientific, industrial, environmental, and medical domains. Unlike carefully curated benchmark datasets, real-world visual data are inherently noisy, heterogeneous, incomplete, heavily imbalanced, and weakly annotated. While recent advances in convolutional neural networks, vision transformers, and multimodal models have achieved remarkable performance on standard benchmarks, their reliability degrades significantly when deployed in real operational settings characterized by scarce labels, sensor noise, or domain shift. Addressing these data-centric constraints is essential to bridge the gap between theoretical model performance and dependable, interpretable, and secure deployment of systems on complex visual data.The primary goal of this Research Topic is to advance data science and machine learning approaches that make computer vision systems robust, reliable, and interpretable when processing imperfect real-world visual big data. Rather than prioritizing marginal improvements in benchmark accuracy, this collection emphasizes data-centric challenges: handling extreme class imbalance, learning from noisy or weakly supervised labels, adapting to distribution shifts, and ensuring reproducibility and uncertainty quantification in domain-specific environments. By bringing together methodological innovations with substantive applications—spanning document analysis, astronomical imaging, biomedical diagnostics, and environmental sensing—this topic aims to establish principled standards for deploying robust vision systems outside the laboratory.We welcome submissions that address robustness in computer vision on real-world visual big data. Submissions should demonstrate a clear methodological contribution, practical significance, or both; studies that solely present benchmark performance comparisons without substantive data science contributions will not be considered.Themes of interest include, but are not limited to:• Deep learning for document, floor-plan, and structured visual data under scarce or noisy annotations• Object detection, segmentation, and anomaly detection with extreme class imbalance• Astronomical and scientific observational big data processing under rare-event constraints• Biomedical and medical image analysis with limited, heterogeneous, or weakly labeled data• Video understanding, vision-language learning, and multimodal fusion under distribution shift• Robustness, explainability, uncertainty estimation, and error analysis in real-world computer vision• Data quality, label noise mitigation, weak supervision, and self-supervised feature representation• Scalable processing frameworks, reproducible benchmarks, and open datasets for constrained visual big data Topic editors: SNIGDHA SEN, Chiranjoy Chattopadhyay, Divya Sharma Source: https://www.frontiersin.org/research-topics/86599/robust-deep-learning-for-real-world-visual-big-data