基于RFB和超网络的跨尺度多层次真实失真图像质量评价方法  

Multi⁃scale and multi⁃level authentic distorted image quality assessment based on RFB and hyper networks

作　　者：周怀博 贾惠珍 王同罕 ZHOU Huaibo;JIA Huizhen;WANG Tonghan(School of Information Engineering,East China University of Technology,Nanchang 330013,China;Jiangxi Engineering Laboratory on Radioactive Geoscience and Big Data Technology,East China University of Technology,Nanchang 330013,China)

机构地区：[1]东华理工大学信息工程学院,江西南昌330013 [2]东华理工大学江西省放射性地学大数据技术工程实验室,江西南昌330013

出　　处：《现代电子技术》2024年第9期47-52,共6页Modern Electronics Technique

基　　金：国家自然科学基金项目(62266001);国家自然科学基金项目(62261001)。

摘　　要：为了能在真实失真图像质量领域实现高效的跨尺度学习,提出一种双分支特征提取方法。首先,利用对比学习方法自监督地提取跨尺度、跨颜色空间的图像内容感知特征;随后,采用基于扩张感受野和超网络的策略,将多层次特征信息与跨尺度信息进行循环交互融合,以获取更贴近人类感知的图像质量特征。基于公开真实失真数据库的实验结果表明,所提算法在真实失真图像质量评价上取得了优越性能,而且,通过两个尺度的实验结果展示了该算法实现了更高效的跨尺度学习,从而为图像多尺度深度网络的应用提供了较好基础。An innovative dual⁃branch feature extraction method is proposed to achieve efficient cross⁃scale learning in the domain of authentic distorted image quality assessment.The method undergoes a two⁃phase training process.In the first phase,cross⁃scale and cross⁃color⁃space image content perception feature is extracted by a self⁃supervised contrast learning approach.In the second phase,a strategy based on dilated receptive fields and hypernetworks is employed to establish a cyclic feature fusion,which circularly interacts and integrates multi⁃level feature information with cross⁃scale information to obtain image quality features closer to human perception.On the basis of the validation on the publicly available authentic distorted image databases,the experimental results demonstrate that the proposed algorithm has achieved superior performance in the quality assessment of authentic distorted images.The experimental results show that the proposed algorithm can realize more efficient cross⁃scale learning,which provides a good foundation for the application of multi⁃scale deep network of image processing.

关 键 词：图像质量评价 无参考 真实失真 跨尺度学习 多特征融合 双分支特征提取 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]