Perceptual quality assessment for no-reference image via optimization-based meta-learning.

[Display omitted] • We integrate the image quality assessment with meta-learning strategy. • This model optimizes initialization parameter and learning rate simultaneously. • Our framework can learn various distortion knowledge and complex real distortions. Image quality assessment (IQA) is a critical issue in computer vision, which intends to simulate human visual system (HVS) with a view to evaluating the error degree of distorted images from reference images. The algorithms based on deep learning have been successfully introduced to quality evaluation of no-referenced images in the last years. Unfortunately, they are confronted with the problems of over-fitting and weak generalization ability caused by insufficient labeled data. The emergence of meta-learning has brought new ideas, which have been proved to address the issues regarding few-shot learning. However, the commonly used meta-learning metrics only learn the initialization of weights, which can't guarantee the optimal gradient direction. The manual design mode leads to lower accuracy and speed. In our work, an improved meta-learning framework is applied to no-reference (NR) IQA to meet the above challenges. It can achieve maximum generalization performance through only a few update iterations. Specifically, we collected a great many NR-IQA tasks with different distortions to pre-train meta-model and optimizer to learn a general weight initialization and optimization rule. Then, the meta-model acquired meta-knowledge and learned unique learning rate for each task. Finally, it could be directly adapted to new NR-IQA tasks only by fine-tuning a few images. Experiments on synthetic and authentic datasets proved that our approach has more vital learning ability and better generalization performance for evaluating real distorted images, effectively reducing dependence on manual marking. [ABSTRACT FROM AUTHOR]