基于多尺度特征融合的无参考图像质量评估  

No-reference Image Quality Assessment Based on Multi-scale Feature Fusion

作　　者：奥宁宁 王贺 AO Ningning;WANG He(College of Physics and Electronic Engineering,Shanxi University,Taiyuan,030006,China)

机构地区：[1]山西大学物理电子工程学院,山西太原030006

出　　处：《网络新媒体技术》2024年第6期31-37,共7页Network New Media Technology

基　　金：2022年山西省高校科技创新项目(编号:2022L008)。

摘　　要：图像质量评估是图像处理领域基础技术之一。无参考图像质量评估模型(NR-IQA)无需参考图像,具有更广泛的适用性,一直是图像质量评估领域研究的热点。本文提出一种基于密集特征金字塔网络和Swin Transformer的无参考图像质量评估算法。密集特征金字塔网络采用层间残差连接、层间密集连接和特征重加权策略,其自顶向下和自底向上的聚合路径能更有效地聚合多尺度特征。Swin Transformer引入窗口注意力机制,更好地捕获图像中的局部和全局信息,减少了计算复杂度。在TID2013数据集上,本文算法的SROCC和PLCC指标相比原算法分别提升2.6%和1.9%,在KADID-10k数据集上分别提升1.4%和0.9%,提高了无参考图像质量评估的性能。Image quality assessment is one of the basic technologies in the image field of image processing.The reference-free image quality assessment model does not require the use of reference images and has wider applicability.It has always been a hot research topic in the field of image quality assessment.This paper proposes a no-reference image quality assessment algorithm based on dense feature pyramid network and Swin Transformer.The dense feature pyramid network uses inter-layer residual connections,inter-layer dense connections and feature re-weighting strategies,and its top-down and bottom-up aggregation paths can aggregate multi-scale features more effectively.Swin Transformer introduces a window attention mechanism to better capture local and global information in images and reduce computational complexity.On the TID2013 data set,the SROCC and PLCC indicators of this algorithm are improved by 2.6%and 1.9%respectively compared with the original algorithm.On the KADID-10k data set,the SROCC and PLCC indicators are improved by 1.4%and 0.9%respectively,which improves the performance of the no-reference image quality assessment.

关 键 词：深度学习 图像质量评估 特征金字塔 多尺度特征融合 Swin TRANSFORMER 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]