用于单图像超分辨率的全局特征高效融合网络

Global Feature Efficient Fusion Network for Single Image Super Resolution

现有图像超分辨率网络中普遍存在对层间特征利用水平较低的现象,使得在图像重建过程中有细节特征丢失,最终处理结果纹理模糊、图像质量欠佳。为此提出一种用于图像超分辨率的全局特征高效融合网络模型。主体使用对称卷积神经网络实现浅层特征的逐级提取,并结合Transformer完成浅层与深层特征的融合利用。设计的对称自指导残差模块可以在浅层网络实现不同层间特征更具表达性的融合,同时提升网络的特征提取能力;特征互导融合模块可以增强网络对浅层特征与深层特征的融合能力,促进更多的特征信息参与到细图像重建过程。在Set5、Set14、BSD100和Urban100数据集上同近年来的经典网络(HR、CARN、IMDN、MADNet、LBNet)进行性能对比,实验结果表明:所提网络模型在峰值信噪比上有所提升,并在视觉直观对比中取得了较好的图像超分辨率效果,可改善超分辨率图像质量欠佳的问题。

In existing image super resolution networks,the utilization of interlayer features stays at low level,which leads to the loss of detail features in image reconstruction,fuzzy texture and poor image quality in the final processing result.In this paper,a global feature efficient fusion network for image super resolution was proposed.The main body of the network has the symmetric convolutional neural network used to extract shallow features step by step,and Transformer combined to realize both integration and utilization of shallow and deep features.The designed symmetric self-directed residual block can achieve more expressive feature fusion between different layers in shallow network and improve feature extraction ability of the network.The feature mutual-directed fusion block can enhance the fusion ability of shallow feature and deep feature,and promote more feature information to participate in the process of fine image reconstruction.Finally,on the Set5,Set14,BSD100 and Urban100 datasets,having its performance compared with that of the recent years' classical networks(HR,CARN,IMDN,MADNeT and LBNeT) shows that,this network can improve the peak signalto-noise ratio,and achieve better image super resolution effect in the visual comparison,as well as improve the poor quality of the SR image.