摘要
图像超分辨率(Super Resolution,SR)重建是计算机视觉领域中提高图像和视频分辨率的一种重要图像处理技术,针对基于深度学习的图像重建模型层次过多以及梯度传输困难导致训练时间长、重建图像视觉效果不理想的问题,本文提出了一种改进残差块和对抗损失的GAN(Generative Adversarial Networks)图像超分辨率重建模型.首先,在模型结构上,设计剔除多余批规范化操作的残差块并组合成生成模型,将深度卷积网络作为判别模型把控重建图像的训练方向,以减少模型的计算量;然后,在损失函数中,引入Earth-Mover距离设计对抗损失以缓解模型梯度消失的问题,采用L1距离作为重建图像与高分辨率图像相似程度的度量以指导模型权重更新来提高重建视觉效果.在DIV2K、Set5、Set14数据集上的实验结果表明:该模型剔除多余批规范化后的训练时间相比改进前模型减少约14%并有效提高图像的重建效果,结合Earth-Mover距离与L1距离的损失函数有效地缓解了梯度消失的问题.模型相较于双三次插值、SRCNN、VDSR、DSRN模型,提高了对低分辨率图像的超分辨率重建效率和视觉效果.
Image super-resolution (SR) reconstruction is an important image processing technology to improve the resolution of image and video in computer vision. Image reconstruction model based on deep learning has not been satisfactory due to the too many layers involved, the excessively long training time resulting from difficult gradient transmission, and the unsatisfactory reconstructed image. This paper proposes a generative adversarial networks (GAN) image SR reconstruction model with improved residual block and adversarial loss. Firstly, on the model structure, the residual blocks of the excess batch normalization were designed and combined into a generative model, and the deep convolution network was used as the discriminant model to control the training direction of the reconstructed image to reduce the model’s calculation amount. Then, in the loss function, the Earth-Mover distance was designed to alleviate model gradient disappearance. The L1 distance was used as the measure of the degree of similarity between the reconstructed image and the high-resolution image to guide the model weight update to improve the reconstructed visual effect. Experimental results from the DIV2K, Set5, and Set14 datasets demonstrate that compared with the model before improvement, the training time of the proposed model was reduced by about 14% and the image reconstruction effect was effectively improved. For the loss function combined with Earth-Mover distance and L1 distance, gradient disappearance was effectively alleviated. Therefore, the proposed model significantly improved the SR reconstruction efficiency and visual effect of low-resolution images compared with Bicubic, SRCNN, VDSR, and DSRN model.
作者
张杨忆
林泓
管钰华
刘春
ZHANG Yangyi;LIN Hong;GUAN Yuhua;LIU Chun(College of Computer Science and Technology,Wuhan University of Technology,Wuhan 430063,China)
出处
《哈尔滨工业大学学报》
EI
CAS
CSCD
北大核心
2019年第11期128-137,共10页
Journal of Harbin Institute of Technology
