基于渐进式特征增强网络的超分辨率重建算法

为了在图像重建质量和网络参数之间取得较好的平衡，本文提出一种基于渐进式特征增强网络的超分辨率（Super-Resolution，SR）重建算法。该方法主要包含两个模块:浅层信息增强模块和深层信息增强模块。在浅层信息增强模块中，首先利用单层卷积层提取低分辨率（Low-Resolution，LR）图像的浅层信息，再通过我们设计的多尺度注意力块来实现特征的提取和增强。深层信息增强模块先利用残差学习块学习图像的深度信息，然后将得到的深层信息通过设计的多尺度注意力块来获得增强后的深层多尺度信息。最后我们利用跳转连接的方式将首层得到的浅层信息和深层多尺度信息进行像素级相加得到融合特征图，再对其进行上采样操作，得到最终的高分辨率（High-Resolution, HR）图像。实验结果表明，相比于一些主流的深度学习超分辨率方法，本文方法重建得到的图像无论是主观效果还是客观指标，都取得了更好的效果。 

Progressive Feature Enhancement Network for Image Super-Resolution Reconstruction

In order to achieve a better balance between image restoration quality and network parameters, this paper proposes a super-resolution (SR) method based on progressive feature enhancement network. The method mainly consists of two modules: shallow information enhancement module and deep information enhancement module. In the shallow information enhancement module, firstly, the shallow information of low-resolution (LR) image is extracted by single convolution layer, and then we design the multi-scale attention block to achieve feature extraction and enhancement. The deep information enhancement module first uses the residual learning block to learn the deep information of the image and then employs the designed multi-scale attention block to obtain the enhanced deep multi-scale information. Finally, we use a skip-connection to fuse the shallow information obtained at the first layer and the deep multi-scale information at the pixel level and then up-sample the fusion feature maps to obtain the final high-resolution (SR) image. The experimental results show that, compared with some mainstream deep learning SR methods, the image reconstructed by the proposed method has achieved better results in both subjective and objective indicators.