.png){kind=icon}
.png){kind=icon}
Return
Cross dropout based dynamic learning for blind super resolution
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2024.129234.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The current explicit modeling methods of single image super resolution (SISR) have limited utilization of implicit information within images, and are not effective for some complex degradation images. These often result in over-smoothed or visually uncomfortable artifacts. Meanwhile, kernel-based methods are sensitive to minor estimation errors. To alleviate these issues, in this paper we propose across dropout based dynamic network (CDDNet) for multi-degradation blind super resolution. Particularly, our proposed CDDNet models the degradation of low-resolution (LR) images using degradation weights as the global attention and includes a gyroscope structure as the local attention mechanism to further improve performance. It can be seen as performing a sparsity process on the feature map, which means that activation values at random locations are absent or weakened. And it can improve the ability of CDDNet to handle multiple deep degradation. In order to enhance the ability of Generative Adversarial Network (GAN) to handle complex two-stage degradation, we combine a GAN loss, an artif loss and a pixel loss to form an integrate loss suitable for complex blind super-resolution tasks. Experimental results show that our proposed CDDNet is effective, achieving higher reconstruction accuracy and acceptable quality of visual perception, with good results on both synthetic and real-world datasets. Code is available at https://github.com/Aries1213/CDDNet.
Keywords:
Blind image super resolution
Generative adversarial networks
Dropout
Attention mechanism
CDDNet