摘要
提出了一种用于多聚焦图像融合的卷积神经网络(CNN)。与现有的基于CNN的图像融合方法将源图像分解成几个小块,然后使用一个分类器来估计图像块是否聚焦相比,本文方法直接将整个图像转换成一个决策图。像素级回归策略可以充分利用互补信息,解决了聚焦/散焦区域周围模糊程度估计的困难。此外,在图像融合领域,应用环形残差网络(RResNet)模块来提取更多聚焦区域的语义信息。同时,利用结构相似度(SSIM)估计生成的融合图像与参考图像之间的结构相似性以提高融合图像的质量,同时采用边缘保留损失函数来保留源图像中更多的梯度信息。实验结果表明:该方法在主观视觉效果和客观评价方面均优于其他融合算法。
In this paper,a novel convolutional neural network(CNN) for multi-focus image fusion is proposed. Compared with existing image fusion methods based on CNN which decompose the source image into several patches and adopt a classifier to estimate whether the patch is focused or defocused,the method in this paper directly converts the whole image into a decision diagram. The pixel-level regression strategy can make use of the complementary information and address the difficulty of estimating blur level around the focused/defocused region. Furthermore,the ringed residual network(RResNet)block is utilized to extract more semantic information from the focused region in image fusion field. In the meanwhile,the structural similarity index(SSIM)loss is utilized to estimate the structural similarity between the generated fusion image and the ground-truth reference to improve the quality of the fused images,and the edge preservation loss function is applied to preserve more gradient information from source image. Experimental results demonstrate that the proposed method is superior to other fusion algorithms in subjective visual effect and objective assessment.
作者
申铉京
张雪峰
王玉
金玉波
SHEN Xuan-jing;ZHANG Xue-feng;WANG Yu;JIN Yu-bo(College of Computer Science and Technology,Jilin University,Changchun 130012,China;Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education,Jilin University,Changchun 130012,China;Changchun Expert Information Technology Co.,Ltd.,Changchun 130012,China)
出处
《吉林大学学报（工学版）》
EI
CAS
CSCD
北大核心
2022年第8期1857-1864,共8页
Journal of Jilin University:Engineering and Technology Edition
基金
国家重点研发计划项目(2018YFB0804203)
国家自然科学基金区域联合基金项目(U19A2057)
国家自然科学基金面上项目(61876070)
吉林省科技发展计划项目(20190303134SF)。
关键词
多聚焦图像融合
深度学习
像素级回归
卷积神经网络
multi-focus image fusion
deep learning
pixel-level regression
convolutional neural network
