Iterative transfer learning with large unlabeled datasets for no-reference image quality assessment

Abstract

No-reference image quality assessment is crucial for evaluating perceptual quality across diverse image-processing applications. Given the challenge of accruing mean opinion scores for images, utilizing data augmentation and transfer learning is vital for training predictive networks. This paper presents a new iterative transfer learning technique, which helps to transfer knowledge between heterogeneous network architectures, and overcomes the problem of overlearning when training on small datasets. The proposed method used a large amount of unlabeled data during training, improving its ability to handle different image quality conditions. We also presented a two-branch convolutional neural network architecture, which merges multi-scale and multi-level attributes efficiently. This architecture emphasizes both local detail extraction and high-level comprehension, and the result was fast execution time and minimal memory overhead. Empirical results showed that applying iterative transfer learning to train a two-branch convolutional neural network achieved superior real-time performance and at the same time exhibited good performance in spearman's rank order correlation coefficient. Furthermore, the model manifested robustness for the noisy mean opinion score, which is prevalent in available datasets, and during data augmentation processes.