D4.3 Report on the methodology of transparent ML models (short version)

Modern object detection models like R-CNN, SSD, and YOLO lack robustness, domain specificity, and interpretability, making artificial intelligence (AI) decisions opaque and prone to bias. This work enhances machine learning (ML) transparency by developing interpretable models with fairness and accountability, particularly in deep learning for vision and image processing. Key efforts include active learning to reduce training data needs, human-in-the-loop feedback for meaningful AI explanations, and fairness improvements using SMOTE-driven techniques. Object detection explanations leverage YOLOWorld and GradCAM for video, while LIME and SHAP aid audio analysis. By integrating explainable AI (XAI) and human-centric design, this research ensures more transparent, fair, and accountable AI systems for real-world applications.