【摘要】 为了改善图像超分辨率重建的效果,针对很多超分辨率重建方法中忽略了特征通道间相关信息以及网络数据传递中信息丢失问题,提出了一种通道注意力与残差级联超分辨率重构网络。首先,对输入的低分辨率图像进行浅层的特征提取;随后,通过残差级联组提取深层特征,利用注意力模块自适应地对特征通道的权重进行校正,融合节点将残差级联组的输出特征和浅层特征进行级联融合,保证低分辨率图像的有效信息在网络传递中不被丢失;最后对提取到的特征信息进行亚像素重构。在不同基准数据集上的实验结果表明,不论从主观视觉上还是客观指标比较,所提方法都要优于其他方法,在Urban100数据集上4倍超分辨率的PSNR指标提高了0.1 dB,这都表明该网络在图像超分辨率重建方面有不错的性能。更多还原

【Abstract】 Several existing image super-resolution reconstruction methods face challenges owing to information loss between feature channels and during network data transmission. To eliminate this problem,a channel attention and residual concatenation network for image super-resolution is proposed to improve the effect of image super-resolution reconstruction. Initially,shallow feature extraction is performed on the input low-resolution image. Then,the deep features are extracted by the residual concatenation group,and subsequently the attention module is used to adaptively correct the weights of feature channels. The fusion node concatenates,and fuses the shallow and output features of the residual concatenation group to ensure that there is no loss of effective information of the low-resolution image during transmission. Finally,the extracted feature information is reconstructed using a sub-pixel. The experimental results on different benchmark datasets indicates that the proposed method achieves better results in subjective vision and objective index comparison than existing methods. On the Urban100 dataset,the PSNR index of 4 times super-resolution is increased by 0. 1 dB. This indicates that the network performs well in image super-resolution reconstruction.更多还原