Abstract: Aiming at low channel feature correlation utilization in the middle layer and nonlinear mapping of low-resolution images and high-resolution images, a dual regression network based on an efficient second-order attention（ESADRNet） is proposed. The network divides the reconstruction task into two regression networks：the original regression network and the dual regression network. The original regression network firstly uses FReLU as the down-sampling layer of the activation function to extract more efficient spatial context features of the image. Meanwhile, the method based on the multi-level skip connection residual attention group（MLSCRAG） composed of the multi-level skip connection residual（MLSCR） and the efficient second-order channel attention（ESOCA） module and shared source connection（SSC） and sub-pixel convolution constructs a progressive up-sampling network, so that the network can focus on more discriminative feature representation, and has more powerful feature expression and feature-related learning capabilities. Finally, the dual regression network is used to constrain the mapping space, and find the optimal reconstruction function. After comparative experiments on the Set5, Set14, BSD100 and Urban109 datasets, it is proved that the network is superior to other methods in terms of objective quantitative indicators and subjective vision.

Key words: super-resolution reconstruction, attention mechanism, dual regression network, convolutional neural network, deep learning

摘要： 针对中间层通道特征相关性利用率低、低分辨率图像和高分辨率图像函数映射空间非线性的问题，提出了一种基于高效二阶注意力机制的对偶回归网络（ESADRNet）。该网络将重建任务分为两个回归网络：原始回归网络和对偶回归网络。原始回归网络采用FReLU为激活函数的下采样层对图像进行更高效的空间上下文特征提取；基于多级跳跃连接残差块（MLSCR）和高效二阶通道注意力模块（ESOCA）构成的多级跳跃连接残差注意力模块（MLSCRAG）、共享源跳跃连接（SSC）和亚像素卷积构建渐进式上采样网络，使网络专注于更具辨别性的特征表示，具有更强大的特征表达和特征相关学习能力；利用对偶回归网络约束映射空间，寻找最优重建函数。在Set5、Set14、BSD100和Urban109数据集上经过对比实验证明，该网络在客观定量指标和主观视觉方面均优于其他对比方法。

关键词: 超分辨率重建, 注意力机制, 对偶回归网络, 卷积神经网络, 深度学习