An Effective Meta-Learning Network Model for No-Reference Image Quality Assessment

The use of meta-learning has been proven efficient to address the limitations of insufficient data for no-reference image quality assessment (NR-IQA). While meta-learning methods have been developed as training process, the works for appropriate network models were not sufficient, which posed limitations on performance improvement. The goal of this work is to design a suitable network model for meta-learning to enhance NR-IQA performance. The proposed method follows the training process of optimization-based meta-learning for each distortion type. The proposed network model learns efficiently distortion-specific features and adapts easily to unknown distortions. Experimental results show that the proposed network model provides superior performance than the previous NR-IQA methods using meta-learning.