基于元原型网络的无参考图像质量评价

No-reference image quality assessment based on meta prototype network

针对基于深度学习的模型因需要大量标注数据而选择在预训练模型上进行微调,导致面对新任务泛化性不足的问题,提出一种基于元原型网络的无参考图像质量评价算法。利用元原型网络提取相关任务中的元知识形成质量先验模型,在面对未知任务时快速泛化。首先,在不同失真的数据集上利用元学习方法获取各种失真的共享先验知识得到质量先验模型;接着,为了能够更好地捕获各种失真场景共享先验知识,利用元原型单元对图像特征进行重建,以获得更加丰富的先验知识,从而便于后续的质量分数预测过程;最后,在目标任务上对质量先验模型进行微调,以构建质量模型。在CID2013、LIVE challenge和KonIQ-10K三个数据库上的实验结果表明,所提方法具有更好的性能。

The model based on deep learning requires a large amount of annotated data,so it is subjected to fine-tune on the pre-training model,which leads to insufficient generalization when facing new tasks.In view of this,a no-reference IQA(NR-IQA)algorithm based on meta prototype network is proposed.The meta prototype network is used to extract meta-knowledge in relevant tasks to form a quality prior model,which can quickly fulfill generalization when facing unknown tasks.The meta-learning method is used to obtain shared prior knowledge of various distortions on different distortion datasets to form the quality prior model.In order to better capture shared prior knowledge of various distortion scenarios,the meta prototype units are used to reconstruct image features to obtain richer prior knowledge,which facilitates the subsequent process of quality score prediction.The quality prior model is fine-tuned on the object task to construct the quality model.Experimental results on the three databases of CID2013,LIVE challenge and KonIQ-10K show that the proposed method has better performance.