| 摘 要: 真实失真图像的无参考图像质量评价(NO Reference Image Quality Assessment,NRIQA)是图像处理领域的一个具有挑战性的问题,现有的模型难以捕获有效的质量感知特征。为了学到更准确的质量感知特征表示,提出一种基于卷积调制和自注意力机制的 NRIQA网络。网络浅层使用卷积调制捕获图像的局部特征,网络深层通过双支路自注意力特征融合模块与线性注意力捕获图像的全局特征。在6个具有代表性的数据集上进行的实验结果表明,该网络表现优异。其中,在 KADID-10K(Konstanz Artificially Distorted Image Quality Database 10K)和LIVEC(LIVE In the Wild Image Quality Challenge Database)数据集上的斯皮尔曼等级相关系数(Spearman Rankorder Correlation Coeficient,SRCC)分别达到了0,918 和 0.882,优于 DEIQT(Data-Eficient lmage Quality Transformer)和 MUSIQ(Multi-Scale lmage Quality Transformer)等先进的无参考图像质量评价算法,预测结果更准确。 |
| 关键词: 无参考图像质量评价;质量感知特征;局部特征;全局特征;双支路自注意力 |
|
中图分类号: TP391.41
文献标识码: A
|
|
| No-Reference Image Quality Assessment Algorithm Based on Multi-ScaleFeatures |
|
ZHANG Jun, ZHANG Xuande
|
(School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, X'i an 710021, China)
zj1739041770@163.com; zhangxuande@sust.edu.cn
|
| Abstract: No-Reference Image Quality Assessment (NRIQA) of real distorted images is a challenging problem in the field of image processing, as existing models struggle to capture effective quality-aware features. To learn more accurate quality-aware feature representations, this paper proposes an NRIQA network based on convolutional modulation and a sel-f attention mechanism. In the shallow layers of the network, convolutional modulation is employed to capture local image features, while in the deep layers, a dua-l path sel-f attention feature fusion module combined with linear attention is utilized to capture global image features. The results of the experiment on six representative datasets demonstrate the superior performance of the proposed network. Specifically, on the KADID-10K (Konstanz Artificially Distorted Image Quality Database 10K) and LIVEC (LIVE In the Wild Image Quality Challenge Database) datasets, the Spearman Rank-order Correlation Coefficient (SRCC) reaches 0.918 and 0.882, respectively, outperforming state-o-f theart NRIQA algorithms such as DEIQT (Data-Efficient Image Quality Transformer) and MUSIQ (Mult-i Scale Image Quality Transformer), with more accurate prediction results. |
| Keywords: No-Reference Image Quality Assessment (NRIQA); quality-aware features; local features; global features; dua-l path sel-f attention |
