Abstract: Abstract: Blind image super-resolution reconstruction is dedicated to restoring low-resolution images into clear high-resolution images in realistic scenes. Most of the current degradation estimation networks cannot estimate the degradation information of low-resolution images well, as well as the generalization ability of the traditional residual dense blocks in the super-resolution network is poor and cannot adapt to the complex degradation types. To address the above issues, this paper proposes a blind super-resolution reconstruction method based on hybrid degradation estimation and dynamic residual dense blocks. The method firstly proposes a hybrid degradation estimation network and uses it to accurately estimate the degradation information of the image, and then proposes a dynamic residual dense block, which improves the universality of the super-resolution reconstruction network, and finally realizes the accurate prediction and high-quality reconstruction of the complex degradation information through the joint optimization from end to end. Experimental results on multiple datasets of different types show that the proposed method achieves an average increase of 1.51 dB in PSNR and 0.038 in SSIM, along with an average decrease of 0.025 in LPIPS, compared to other competing methods, which proves that the proposed method has better detail recovery capability and artifact suppression effect.

Key words: Keywords: image reconstruction, blind image super-resolution, generative adversarial network, hybrid degenerate estimation network, dynamic residual dense blocks

摘要： 摘 要: 盲图像超分辨率重建致力于在真实场景下将低分辨率图像恢复成清晰的高分辨率图像。目前大多数退化估计网络不能较好地估计低分辨率图像的退化信息，以及超分辨率网络中传统残差密集块的泛化能力较差，无法适应复杂的退化类型。针对以上问题本文提出了混合退化估计与动态残差密集块的盲超分辨率重建方法。该方法首先提出一种混合退化估计网络，并将其用于对图像退化信息进行精确估计，然后提出了一种动态残差密集块，提高了超分辨率重建网络的普适性，最后通过端到端的联合优化实现了对复杂退化信息的精准预测与高质量重建。在多个不同类型数据集上的实验结果表明，所提方法相比其他对比方法的PSNR平均提高了1.51dB，SSIM平均提高了0.038，LPIPS值平均降低了?0.025，证明所提方法具有更优的细节恢复能力与伪影抑制效果。

关键词: 关键词: 图像重建, 盲图像超分辨率, 生成对抗网络, 混合退化估计网络, 动态残差密集块