基于深度学习的单幅图像超分辨率重建方法研究

Research on Super-Resolution Reconstruction of Single Image Based on Depth Learning

为了解决基于单幅图像自适应稠密连接超分辨率(ADCSR)算法中的残差单元的融合问题,本文提出了一种基于行稀疏约束l_(0,2)-范数和soft-max运算的新策略.根据ADCSR算法,本文算法分为两部分:BODY和SKIP,前者专注图像的高频特征学习,后者专注低频特征学习.BODY部分中所有自适应密集残差单元(ADRU)的输出,作为初始特征图,可用特征数目l_(0,2)-范数作为活动水平度量,然后利用基于块的平均算子计算最终活动水平图,最后利用soft-max得到融合后特征映射,改进了原ADCSR算法中卷积融合粗糙的缺点,保留了更多的结构信息和特征.此外特征数目l_(0,2)-范数作为字典原子更加精确地获取更高的权重,获得了更优的峰值信噪比PSNR、结构相似性SSIM和视觉效果,计算机实验证明了本文算法的有效性.

In order to solve the fusion problem of residual elements in ADCSR(adaptive dense connection super-resolution)super-resolution algorithm based on single image,a new strategy based on row sparse constraint l_(0,2)-norm and soft-max operation is proposed.According to the ADCSR algorithm,this algorithm is divided into two parts:BODY and SKIP.The former focuses on the high-frequency feature learning of the image,while the latter focuses on low-frequency feature learning.The outputs of all ADRUs(adaptive dense residual units)in the BODY part are used as the initial feature map,and the number of feature l_(0,2)-norm can be used as the activity level measurement.Then the block based average operator is used to calculate the final activity level map,and finally the fused feature map is obtained by soft-max,which improves the disadvantage of rough convolution fusion in the original ADCSR algorithmand retains more structure information and features.In addition,the number of features l_(0,2)-norm as a dictionary atom can obtain higher weight more accurately,and obtain better PSNR,SSIM and visual effect.Computer experiments show the effectiveness of the algorithm in this paper.