结合误差反馈的LapSRN的图像超分辨率重建

Image super-resolution via Laplacian pyramid network with error feedback mechanism

近年来,神经网络被广泛应用于超分辨率重建,但是多数网络都是单一学习低分辨率(LR)图像和高分辨率(HR)图像之间的映射关系,而文中提出了一种基于拉普拉斯金字塔网络(LapSRN)的算法,考虑LR图像和不同尺寸HR图像之间的映射,挖掘新的特征信息。另外,针对网络中多次连续上采样所带来的误差,引入错误反馈机制,将之前的采样误差传递给当前采样层,这样不断地进行迭代和自校正,产生更优解。实验表明,所提方法比现有的相关方法在PSNR指标上有普遍提高,细节恢复效果和收敛速度也有明显提升。

In recent years,neural networks have been widely used in the resolution of the reconstruction,but most of the network study is a single Low Resolution(LR)image and the mapping relationship between High Resolution(HR)image,and this paper puts forward a based on the Laplacian pyramid network(LapSRN)algorithm,considering the LR image and mapping between different size HR image,digging a new feature information.In addition,an error feedback mechanism is introduced to solve the errors caused by the continuous up-sampling in the network,and the previous sampling errors are transferred to the current sampling layer,so as to continuously iterate and self-correct and produce better solutions.Experimental results show that the proposed method is generally better than the existing method in PSNR index,and the detail recovery effect and convergence speed are also significantly improved.