【摘要】 研究单图像超分辨率卷积神经网络中递归、反馈和注意力机制等设计，提出了一种多反馈稀疏网络（multi-feedback sparsity network,MFSN）。首先，递归和反馈机制旨在反馈高级特征以增强低级特征。MFSN提出了一种多反馈递归机制，网络中的所有基本模块都可进行多次迭代，每次迭代过程中又有多个位置来反馈特征，丰富了传统递归反馈的形式。其次，研究了稀疏性作用。与注意力机制类似，稀疏性思想也旨在关注图像中的重要信息，但此处的稀疏性与损失函数相关联。更多还原

【Abstract】 Targeting at the exploration of some designs such as recurrent, feedback and attention mechanism, we propose our multi-feedback sparsity network(MFSN). Firstly, for the recurrent and feedback mechanism, it means feeding high-level features back to enhance low-level features. In our MFSN, based on the recurrent and feedback mechanism, a multi-feedback recurrent mechanism is proposed. Specifically, we make all the basic modules iterative in our network. In each iteration, multiple feedback branches corresponding to different positions are fed back. Such a multi-feedback recurrent mechanism enriches the form of recurrent feedback and makes the low-level features get further enhancement. Secondly, we study the effect of sparsity on improving network performance. Similar to the attention mechanism, the sparsity idea also aims to focus on the important information in the image but the sparsity is related to the loss function. Thus, we perform a comparison between them and resulting in the sparsity idea performing slightly better. Finally, a few ablation studies have been executed to validate the effect of our designs. Then we trained our model using a remote sensing dataset to validate its performance in remote sensing. We also compared our model with some models with similar model sizes and got competitive results.更多还原