跨失真表征的特征聚合无参考图像质量评价

Cross-distortion representation of feature aggregation for no-reference image quality assessment

针对现有的基于深度学习的无参考图像质量评价模型容易过拟合,并且对真实失真场景中的未知失真类型难辨识问题,提出一个跨失真表征的特征聚合无参考图像质量评价框架。采用与模型无关的元学习优化算法,学习跨不同失真类型的特征表达,缓解模型过拟合影响;将元学习算法与注意力机制、图神经网络相结合,设计特征聚合模块学习每种失真类型的显著性特征;设计图表示模块学习每种失真类型共有的失真信息,削弱图像内容变化对质量预测的影响。实验结果表明,所提算法在预测真实失真图像质量时能够充分挖掘失真图像的高级语义信息,有效解决真实失真场景下失真图像内容变化、未知失真类型复杂的问题,具有较强的推理和泛化能力。

To solve the problem that the current no-referenced image quality assessment(NR-IQA)models based on deep lear-ning are easy to overfit and unable to adapt to unknown distortions in authentic distortion scenes,a cross-distortion representation of feature aggregation for no-reference image quality assessment(CDR-FA NR-IQA)architecture was proposed.The CDR-FA network adopted a model-agnostic meta-learning(MAML)optimization algorithm to learn shared quality prior knowledge among different distortions and further to alleviate the influence of overfitted model.Combining meta-learning algorithm with attention-based mechanism and graph neural network,a feature aggregation module was designed to learn the prominent features of each distortion,and a graph representation module was designed to learn the common quality prior knowledge between the same distortion to reduce the impact of great content variation on quality prediction.The experiments verify that this algorithm can comprehensively exploit the high-level semantic features of distorted images.Furthermore,the model can effectively solve the great content variation and distortion diversity in the actually distorted scenes,and has strong generalization ability.