基于元学习和失真感知的图像质量评价  

Image quality assessment based on meta-learning and distortion perception

作　　者：万丙辰 张选德 WAN Bingchen;ZHANG Xuande(School of Electronic Information and Artificial Intelligence,Shaanxi University of Science&Technology,Xi'an 710021,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,陕西西安710021

出　　处：《液晶与显示》2024年第11期1519-1531,共13页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.61871260)。

摘　　要：获取图像主观质量评分的成本较高,使图像质量评价(Image Quality Assessment,IQA)模型通常面临训练样本量不足的问题,另外,失真类型对于图像视觉感知质量具有重要的影响。针对以上问题,本文提出了一种基于元学习和失真感知相结合的图像质量评价方法。首先,通过元学习模拟人类学习的过程来快速获取已知失真类型的先验知识,指导后续的ResNet-50网络有效融合多尺度特征。引入失真感知模块捕获完整失真信息,建立统一的质量评价体系。在LIVE、KonIQ-10k等合成失真与真实失真数据集上的实验结果表明,所提模型在小样本条件下,能够提升模型在不同失真类型间的泛化性能。在与现有先进方法的综合实验对比中,本模型较次优方法在PLCC和SROCC两个评价指标上分别取得了1.02%和1.85%的提升。本文模型的评价精度与目前主流的IQA模型相比,具有一定的竞争力。The cost of obtaining subjective quality scores for images is often high,image quality assessment(IQA)models commonly face the challenge of insufficient training samples.Additionally,the type of distortion has a significant impact on the perceived visual quality of images.In light of these considerations,this paper proposes an image quality assessment method that combines meta-learning and distortion perception.Firstly,meta-learning is employed to simulate the human learning process,enabling the rapid acquisition of prior knowledge about known distortion types.This knowledge guides the subsequent ResNet-50 network in effectively integrating multi-scale features.The introduction of a distortion perception module captures comprehensive distortion information,establishing a unified quality assessment framework.Experimental results on synthetic and real distortion datasets such as LIVE and KonIQ-10k indicate that the proposed model under small sample conditions can enhance the generalization performance across different distortion types.In comparison with the existing advanced methods,the model in this paper achieves 1.02%and 1.85%improvement in PLCC and SROCC evaluation indexes compared with the second-best method.In comparison with mainstream IQA models,the evaluation accuracy shows competitive performance.

关 键 词：无参考图像质量评价 元学习 失真感知 泛化性能 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]