The Effect of Uncertainty on No-Reference Image Quality Assessment

No-reference Image Quality Assessment (NR-IQA) attained acceptable results through deep learning models. However, the overfitting, caused by complex deep models and insufficient labeled datasets, has become a primary challenge for the research community. Addressing this issue, various strategies such as data augmentation, transfer learning, and weakly supervised learning have been investigated. This paper introduces an approach, suggesting the use of a probability distribution instead of a rigid target to mitigate overconfidence issues. The proposed label uncertainty can provide acceptable results, especially in terms of cross-dataset validation.