基于Dirac残差模块的单幅图像超分辨率重建

针对单幅图像超分辨率重建问题(SISR),提出了一种新的基于Dirac残差的超分辨率重建算法.算法使用全局跳跃重建层来直接利用输入LR图像的低频特征,通过多个dirac残差块来自适应学习输入LR图像的高频特征,通过亚像素卷积进行图像重建.算法通过权重参数化来改进残差层,同时使用输入图像的卷积特征与残差网络学习特征结合进行重建.实验采用Adam优化器进行网络训练.使用L1范数作为损失函数.在PSNR、SSIM和视觉效果与其他先进算法进行对比,实验结果表明,在常用测试集上与其他深度学习算法相比有较大提高.

Image super-resolution reconstruction based on Dirac residual

For a better performance in single image super-resolution(SISR), we present a new image super-resolution method based on Dirac residual learning. The algorithm directly utilizes the low frequency features of the input LR image through a global jump reconstruction layer. And its high frequency features are learned by a plurality of Dirac residual blocks. Finally, the image reconstruction is performed by sub-pixel convolution. Residual layers are optimized by weight parameterization. The reconstruction combines the convolutional features of the input image with the features of residual learning. We train our model with ADAM optimizer and employ L1 as the loss function. Extensive qualitative and quantitative evaluation results on benchmark datasets verify the great improvement of the proposed algorithm compared with other state-of-the-art SISR approaches.