基于对偶学习策略的单图像超分辨率重建网络

Single-image super-resolution network based on dual learning strategy

针对单图像低分辨率到高分辨率映射具有不适定性、特征图空间信息利用率低下以及网络参数量过大的问题,提出了一种基于渐进上采样的对偶学习算法用于图像的超分辨率重建。首先采用深度可分离卷积使得模型参数量显著减少;再基于亚像素卷积构建渐进上采样网络来高效利用特征图上下文信息;最后利用对偶学习策略构建闭环反馈网络,通过对偶关系相互约束映射空间以获取最佳重建函数。在Set5、Set14、BSDS100、Urban100、Manga109基准数据集上与其他主流的超分辨率方法相比,该算法表现出更优越的性能:有效减少了网络9%的参数量,在×4、×8放大因子下能重建出更清晰的图像,同时能有效缓解图像边缘失真和伪影现象,并且×8放大时的平均峰值信噪比和结构相似度(PSNR/SSIM)分别为26.90/0.751、24.84/0.645、24.74/0.619、22.30/0.560、24.38/0.706。

Aiming at the ill-posedness of the single-image low-resolution to super-resolution mapping,the low utilization of feature space information by the super-resolution reconstruction network,and the excessive amount of network parameters,this paper proposed a dual learning algorithm based on progressive up-sampling,which was applied to super-resolution reconstruction of a single image.The algorithm adopted deep separable convolution to significantly reduced the amount of network parameters,and constructed the progressive up-sampling network based on sub-pixel convolution to efficiently use the spatial information of the feature image.Meanwhile,it used dual learning to construct a closed-loop feedback connection network to obtain the optimal mapping function to estimate the down-sampling kernel to reconstruct low-resolution images.Analyzing on benchmark datasets such as Set5,Set14,BSDS100,Urban100,and Manga109 compared with the state-of-the-arts models,this algorithm can reduce the number of parameters by 9%,and can effectively alleviate the image edge distortion and artifact phenomenon under the large factors,the average PSRN/SSIM are 26.90/0.751、24.84/0.645、24.74/0.619、22.30/0.560、24.38/0.706.