The convolutional neural network plays an important role in single-image super-resolution reconstruction. In this study, we presented a branch residual pyramid channel network to reconstruct high-resolution image from single low-resolution image. Specifically, the low-level information was reconstructed based on the low-resolution image features. Then, we completed the characteristics of the advanced information mapping of the residual neural network. A different number of serial sequence and cross-combined convolution repeatedly were applied to strengthen the key information of mining capacity which increased the multi-branch structure while kept the direct path, we used. In addition, a feature pyramid channel attention module (feature maps aggregated through a top–down path) in each horizontal connection was utilized which generate more targeted feature maps. Experimental results with extensive quantitative and qualitative evaluation on benchmark data sets demonstrated that the proposed algorithm is not only superior to the existing advanced algorithms in speed and accuracy, but also shows good effect in super-resolution reconstruction of single image. Specifically, performance of the proposed method is particularly outstanding on SET5 data set (PSNR value is 0.08 higher than that of the Subprime network), SET14 data set (PSNR value is 0.03 higher than that of the Subprime network) and Urban100 data set (PSNR value is 0.15 higher than that of the Subprime network) when the scaling factor is 2, 3 and 4, respectively.

卷积神经网络在单幅图像超分辨率重建中发挥着重要作用。在这项研究中，我们提出了一个分支残差金字塔通道网络来从单个低分辨率图像重建高分辨率图像。具体来说，低层信息是基于低分辨率图像特征重建的。然后，我们完成了残差神经网络的高级信息映射的特点。我们使用了不同数量的串行序列和交叉组合卷积来加强挖掘能力的关键信息，在保持直接路径的同时增加了多分支结构。此外，每个水平连接中的特征金字塔通道注意模块（通过自上而下的路径聚合的特征图）被利用，以生成更有针对性的特征图。对基准数据集进行广泛的定量和定性评估的实验结果表明，该算法不仅在速度和精度上优于现有的先进算法，而且在单幅图像的超分辨率重建中也显示出良好的效果。具体而言，该方法在SET5数据集（PSNR值比Subprime网络高0.08）、SET14数据集（PSNR值比Subprime网络高0.03）和Urban100数据集（PSNR值高0.03）上的性能尤为突出。当比例因子分别为 2、3 和 4 时，该值比 Subprime 网络的值高 0.15）。