Summary

Visual models based on deep learning have been widely applied in tasks such as image generation and recognition. However, these models are vulnerable to perturbations that do not affect human perception, such as noise, weather effects, and blur, leading to inaccurate predictions. Additionally, adversarial attacks and out-of-distribution data further challenge model robustness. To address these issues, Recommendation ITU-T F.748.82 presents a framework aimed at enhancing the reliability and accuracy of visual models during the training process, ensuring they produce correct predictions under various conditions.