【摘要】 提出一种用于单图像超分辨率任务的多尺度特征融合网络，该网络包含特征下采样、多尺度特征融合和增加额外约束方案。首先，构建了包含Pixel-Unshuffle操作的特征下采样模块获得浅层多尺度特征，进一步学习深层多尺度特征融合，并通过多尺度特征融合提高超分辨率性能。此外，使用包含Pixel-Unshuffle操作的特征下采样模块从SR图像中重建LR图像，并计算其与LR图像之间的损失增加额外约束。大量的实验证明，提出的方法与现有的方法相比表现出具有竞争力的性能。更多还原

【Abstract】 A multi-resolution feature fusion network with additional constraint for image SR is provided in this paper. The network contains feature downsampling and multi-resolution feature fusion as well as an augmentation constraint scheme. Specifically,we obtain multi-resolution features through pixel-unshuffle downsampling block(PDB), then fuse these multi-resolution features with the deep features. We utilize the multi-resolution feature fusion(MRF)module to improve image SR performance. In addition,we use the pixel-unshuffle downsampling block(PDB)module to estimate LR images from SR images, which provide an extra constraint on SR model. Extensive experiments demonstrate that our method achieves competitive performance against existing methods.更多还原