[1]林重成,王同罕*,贾惠珍.基于大核卷积分解和多尺度注意力的图像质量评价[J].计算机技术与发展,2025,(06):175-181.[doi:10.20165/j.cnki.ISSN1673-629X.2025.0011]
　LIN Chong-cheng,WANG Tong-han*,JIA Hui-zhen.Image Quality Assessment Based on Decomposed Large Kernel Convolution and Multi-scale Attention[J].,2025,(06):175-181.[doi:10.20165/j.cnki.ISSN1673-629X.2025.0011]
点击复制

基于大核卷积分解和多尺度注意力的图像质量评价()

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
期数:
2025年06期
页码:
175-181
栏目:
人工智能
出版日期:
2025-06-10

文章信息/Info

Title:
Image Quality Assessment Based on Decomposed Large Kernel Convolution and Multi-scale Attention
文章编号:
1673-629X(2025)06-0175-07
作者:
林重成; 王同罕*; 贾惠珍
东华理工大学 信息工程学院,江西 南昌 330013
Author(s):
LIN Chong-cheng; WANG Tong-han*; JIA Hui-zhen
School of Information Engineering,East China University of Technology,Nanchang 330013,China
关键词:
大核卷积分解; 无参考图像质量评价; 自然场景统计; 多尺度注意力机制; 深度学习
Keywords:
large - kernel convolutional decomposition;  no - reference image quality assessment;  natural scene statistics;  multi - scale attention mechanism; deep learning
分类号:
TP391
DOI:
10.20165/j.cnki.ISSN1673-629X.2025.0011
摘要:
近年来,计算机视觉领域广泛采用注意力机制来模拟人类视觉系统(HVS)的感知过程。 然而,自注意力模型(如 Transformer)在图像处理中的应用面临计算复杂度高和信息丢失的挑战。 为此,该文提出了一种基于大核卷积分解和多尺度注意力机制的图像质量评价方法(LMA-BIQA)。 该方法采用双分支结构网络设计,其中一个分支基于注意力机制,通过大核卷积捕获图像中远距离特征之间的依赖关系,以生成多尺度注意力特征,并通过分解大核卷积实现模型的轻量化,用于提取图像的高级内容特征。 另一个分支利用自然场景统计(NSS)方法提取图像的低级质量特征,以补偿大核卷积在裁剪缩放过程中导致的信息丢失。 该方法在 LIVE、CISQ、TID2013 和 LIVE-C 四个标准数据集上均表现出优异的性能,加权平均后的斯皮尔曼等级相关系数和皮尔森线性相关系数分别达到 0. 928 5 和 0. 937 5。 此外,LMA-BIQA 显著减少了计算时间,参数量和计算时间对比 Re-IQA 模型降低了 65% 以上,且与人类主观评价有较高的一致性。
Abstract:
In recent years,attention mechanisms have been widely used in the field of computer vision to simulate the perceptual process of the human visual system (HVS). However,the application of self-attention models,such as Transformers,in image processing faces challenges of high computational complexity and information loss. To address this, an image quality assessment method based on decomposed large kernel convolution and multi-scale attention mechanisms (LMA-BIQA) is proposed. This method employs a dual-branch network design. One of the branches is based on the attention mechanism. Through large kernel convolution,the dependency rela-tionships between long-range features in the image are captured to generate multi-scale attention features. The lightweight of the model is achieved by decomposing large kernel convolution,which is used to extract advanced content features of the image. Another branch uses the Natural Scene Statistics (NSS) method to extract the low-level quality features of the image to compensate for the information loss caused by large kernel convolution during the cropping and scaling process. The proposed method demonstrates excellent performance across four standard datasets: LIVE, CISQ, TID2013, and LIVE - C. The weighted average Spearman rank correlation coefficient and Pearson linear correlation coefficient reach 0. 928 5 and 0. 937 5,respectively. Additionally,LMA-BIQA significantly reduces computation time,with parameter count and computational time reduced by over 65% compared to the Re-IQA model,while maintaining a high consistency with human subjective evaluations.

相似文献/References:

[1]周姣姣,吴亚东. 基于Curvelet变换的无参考图像质量评价[J].计算机技术与发展,2015,25(07):86.
　ZHOU Jiao-jiao,WU Ya-dong. No-reference Image Quality Assessment Based on Curvelet Transform[J].,2015,25(06):86.
[2]温 静,白 鑫.自适应融合局部和全局特征的图像质量评价[J].计算机技术与发展,2022,32(11):50.[doi:10. 3969 / j. issn. 1673-629X. 2022. 11. 008]
　WEN Jing,BAI Xin.Adaptive Fusion of Local and Global Features for Image Quality Assessment[J].,2022,32(06):50.[doi:10. 3969 / j. issn. 1673-629X. 2022. 11. 008]

更新日期/Last Update: 2025-06-10