基于双分支融合的反馈迭代金字塔去模糊和超分辨率算法

针对低分辨率模糊图像实施超分辨率重建后出现大量伪影和边缘纹理不清晰问题,提出了一种双分支融合的反馈迭代金字塔算法.首先采用不同的分支模块分别提取低分辨率模糊图像中潜在的去模糊特征和超分辨率特征信息;然后采用自适应融合机制将两种不同性质的特征进行信息匹配,使网络在去模糊和超分辨率重建模块中更加关注模糊区域;其次使用迭代金字塔重建模块将低分辨率模糊图像渐进重建为逼近真实分布的超分辨率清晰图像;最后重建图像通过分支反馈模块生成清晰低分辨率图像,构建反馈监督.在GOPRO数据集中与现有算法的对比实验结果表明,所提算法能够生成纹理细节更加清晰的超分辨率图像.

Iterative pyramid deblurring and super-resolution network based on dual-branch fusion feedback

To solve the problems of a large number of artifacts and unclear edge textures in super-resolution reconstruction of low-blurred resolution images, this paper proposed an algorithm named iterative pyramid deblurring and super-resolution network based on dual-branch fusion feedback. The implementation steps of the algorithm are as follows. First, different branch modules extracted the potential deblurring features and super-resolution feature information in the low-resolution blurred image. Second, the adaptive fusion mechanism matched the information from two different features to make sure that the network pays more attention to the blurred area in the deblurring and super-resolution reconstruction modules. Then, the iterative pyramid reconstruction module gradually reconstructed the low-resolution blurred image into a super-resolution clear image which was close to the real distribution. Finally, the reconstructed image was fed into the feedback module to generate sharp low-resolution images by feedback supervision. The experimental results of comparison with existing algorithms in the GOPRO dataset show that the proposed algorithm can generate super-resolution images with clearer texture details.