No-Reference Image Quality Assessment using Meta-SGD
Qingqing Yan, Bei Dong, Xin Yan Li, Yiwen Xiang, Dapeng Luo, Longsheng Wei · 2021 China Automation Congress (CAC) · 2021
Deep learning-based no-reference image quality assessment (NR-IQA) algorithm requires massive training data and labels, but data scarcity usually exists in practical applications. Most of the previous metrics used pre-trained networks to solve this problem by fine-tuning on the target tasks. Unfortunately, the network trained on other tasks cannot be directly applied to IQA tasks and failed when encountering complex distortions. To reduce dependence on labeled data and improve generalization ability of unknown distortion types, a NR-IQA metric based on Meta-SGD is proposed, which can learn the quick adaptability of humans and obtain general meta-knowledge for distortion evaluation. Specifically, we collected a great many NR-IQA tasks with different distortion types to pre-train the model; then a meta-learning metric based on optimization, Meta-SGD, is proposed to acquire the meta-knowledge when evaluating images quality with various known distortions; finally, the obtained pre-trained meta-model can be directly applied to the new NR-IQA tasks only by fine-tuning a few images. Experiments on the TID2013 database demonstrate that the model after meta-learning obviously outperforms than some existing methods, and the average SRCC value for all types of distortion performance tests can reach 0.91.