首页 | 官方网站   微博 | 高级检索  
     

基于多尺度特征映射网络的图像超分辨率重建
引用本文:段然,周登文,赵丽娟,柴晓亮.基于多尺度特征映射网络的图像超分辨率重建[J].浙江大学学报(自然科学版 ),2019,53(7):1331-1339.
作者姓名:段然  周登文  赵丽娟  柴晓亮
作者单位:华北电力大学 控制与计算机工程学院,北京 102206
摘    要:针对基于卷积神经网络的图像超分辨率重建(SRCNN)方法存在的重建网络浅、特征利用率低以及重建图像模糊等问题,提出基于多尺度特征映射网络的图像超分辨率重建方法. 多尺度特征映射网络通过学习低分辨率(LR)特征与高分辨率(HR)特征之间的映射关系,将多个尺度的LR特征映射到HR特征空间,通过特征融合来提高重建过程中对特征的利用率;该方法定义了结合逐像素损失、感知损失和对抗损失的联合损失函数,从低频内容、图像边缘和局部纹理等方面均衡提升重建图像质量. 对数据集Set5、Set14和BSD100的图片4倍下采样后进行测试,与当前主流方法进行比较和分析. 实验证明,基于生成对抗的多尺度特征映射网络在提高图像感知质量方面表现优秀,重建的图像具有更加清晰的边缘和纹理,在客观评价上具有较好的评分.

关 键 词:卷积神经网络  超分辨率重建  生成对抗网络  深度学习  感知损失  

Image super-resolution reconstruction based on multi-scale feature mapping network
Ran DUAN,Deng-wen ZHOU,Li-juan ZHAO,Xiao-liang CHAI.Image super-resolution reconstruction based on multi-scale feature mapping network[J].Journal of Zhejiang University(Engineering Science),2019,53(7):1331-1339.
Authors:Ran DUAN  Deng-wen ZHOU  Li-juan ZHAO  Xiao-liang CHAI
Abstract:An image super-resolution reconstruction method based on multi-scale feature mapping network was proposed for the problems of shallow network, low utilization rate of features and fuzzy reconstructed images, which existed in the super-resolution convolutional neural network (SRCNN). Multi-scale low-resolution (LR) features were mapped into high-resolution (HR) feature space by learning the mapping relation between LR features and HR features, and the utilization rate of features in the reconstruction process was improved by using feature concatenation. A joint loss function consisting of the pixel-wise loss, the perceptual loss and the adversarial loss was defined, which performed well in restoring the low-frequency content, the sharp edges and the high-frequency textures of the reconstructed images. The experimental results of datasets Set5, Set14 and BSD100 for the upscaling factor 4 were compared with those of state-of-the-art methods. The proposed method performs well in improving the perceptual quality of the reconstructed images in order to achieve clearer edges and textures, and has better scores in the objective evaluation.
Keywords:convolutional neural network  super-resolution reconstruction  generative adversarial network  deep learning  perceptual loss  
本文献已被 CNKI 等数据库收录!
点击此处可从《浙江大学学报(自然科学版 )》浏览原始摘要信息
点击此处可从《浙江大学学报(自然科学版 )》下载全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司    京ICP备09084417号-23

京公网安备 11010802026262号