International Journal of Knowledge Processing Studies

Abstract The growing demand for multilingual text and script detection in video scenes across media, education, and surveillance requires accurate real-time recognition despite challenges such as dynamic environments, diverse scripts, low-resource languages, complex layouts, and ethical concerns. This review compares CNNs, RNNs, transformers, and hybrid models in terms of performance, advantages, limitations, and applications. A systematic review of studies published from 2012 to 2024 was conducted using IEEE Xplore, ACM Digital Library, and Google Scholar. Studies were included based on relevance to deep learning-based multilingual text detection, text recognition, or script recognition in video-scene environments and the availability of quantitative evaluation results. Findings highlight the potential of hybrid architectures, particularly Mask TextSpotter v3, for end-to-end text spotting. However, cross-study comparison is limited by variations in datasets, evaluation protocols, and reporting practices. A total of 58 studies met the eligibility criteria. Key trends include cross-lingual transfer learning, multimodal integration, and contextual understanding, while future directions include self-supervised learning, privacy-aware recognition, and NLP-integrated systems.

Abstract The rapid growth of social media and generative artificial intelligence has increased the prevalence of misinformation and manipulated digital media, creating challenges for reliable content verification. This study presents a soft computing-based explainable multimodal framework for digital media integrity risk assessment under uncertainty. The framework integrates transformer-based textual credibility and emotion analysis, OCR-assisted text extraction, CLIP-based image–text semantic consistency, and ResNet50-based visual integrity assessment. The resulting evidence is integrated through a weight-adjusted Mamdani fuzzy inference mechanism to produce continuous integrity-risk scores and graded risk levels. SHAP and Grad-CAM provide supporting textual and visual explanations. On the benchmark test set, the framework achieved 95.6% accuracy and 95.1% F1-score, with a ROC-AUC of 0.97, indicating strong threshold-independent discrimination under the evaluated test conditions. Ablation and sensitivity analyses demonstrated performance changes associated with the integrated components and fusion configuration. Five repeated runs showed limited performance variation under the evaluated protocol. External validation on approximately 1,500 previously unseen multimodal samples achieved a 92.9% F1-score, indicating performance variation under a different data distribution. The framework provides an interpretable, uncertainty-aware approach to multimodal digital media integrity assessment and supports graded risk interpretation rather than relying solely on binary classification.

Abstract This study maps expert-perceived interdependencies and priorities among factors associated with knowledge acquisition in Iran’s marine sector, where distributed operations, personnel mobility, and safety-critical tasks make systematic knowledge capture strategically important. Using a hybrid DEMATEL-based analytic network process (DANP), the analysis draws on 20 individual direct-influence matrices supplied by experts in marine operations, safety management, ship operations, and academia. The criterion framework links organizational capabilities as opportunity conditions, the theory of planned behavior as a motivational gateway, and SECI as a knowledge-conversion lens across six clusters and 20 sub-criteria. Behavioral factors receive the highest cluster priority, closely followed by organizational culture, information technology, and learning conditions. In the aggregated expert-judgment network, perceived behavioral control, attitude, subjective norms, recognition of knowledge value, management practices, strategic alignment, and IT infrastructure occupy cause-side positions, whereas supportive culture is high-priority but effect-side. Reward priorities are close: personal growth and reputation have near-identical point estimates and both rank above cash rewards within this expert panel, but the small gaps do not establish a universal reward hierarchy. Priority rankings remain highly consistent under median and 10% trimmed-mean aggregation; median aggregation changes the cause-side/effect-side classification of only three sub-criteria with near-neutral directional positions, whereas all directional classifications remain unchanged under leave-one-expert-out analysis. Experts show stronger concordance for outgoing than incoming influence rankings. The study distinguishes theoretical relevance, network priority, and directional position, and supports context-sensitive pilot interventions rather than universal causal prescriptions.