RepSR: Training Efficient VGG-style Super-Resolution Networks with Structural Re-Parameterization and Batch Normalization
(RepSR:通过结构重新参数化和批量归一化来训练高效的VGG式超分辨率网络)
作者:1.Xintao Wang, 2.Chao Dong,3.Ying Shan
单位:ARC Lab, Tencent PCG, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences; Shanghai AI Laboratory
代码:https://github.com/TencentARC/RepSR.(暂未放出)
论文地址:https://arxiv.org/pdf/2205.05671
一、问题亮点
1.将训练的多分支架构sr网络模型直接用于测试通常会导致较差的性能表现,要如何去解决这个问题?
2.在当前的sr网络中BN通常被替换或删去,应为他通常会带来性能损失和奇怪的伪影,还有就是bn在sr领域也缺乏相关研究。
<