董彪, 张生, 韩韧. 图像超分辨率重建的循环两级残差网络[J]. 信息与控制, 2023, 52(5): 669-678, 688. DOI: 10.13976/j.cnki.xk.2023.2354
引用本文: 董彪, 张生, 韩韧. 图像超分辨率重建的循环两级残差网络[J]. 信息与控制, 2023, 52(5): 669-678, 688. DOI: 10.13976/j.cnki.xk.2023.2354
DONG Biao, ZHANG Sheng, HAN Ren. Recurrent Dual-level Residual Network for Image Super-resolution Reconstruction[J]. INFORMATION AND CONTROL, 2023, 52(5): 669-678, 688. DOI: 10.13976/j.cnki.xk.2023.2354
Citation: DONG Biao, ZHANG Sheng, HAN Ren. Recurrent Dual-level Residual Network for Image Super-resolution Reconstruction[J]. INFORMATION AND CONTROL, 2023, 52(5): 669-678, 688. DOI: 10.13976/j.cnki.xk.2023.2354

图像超分辨率重建的循环两级残差网络

Recurrent Dual-level Residual Network for Image Super-resolution Reconstruction

  • 摘要: 针对当前基于深度学习图像超分辨率重建方法模型规模大、重建效率低等问题,提出了一种能够获得性能和网络规模优越平衡的图像超分辨率重建网络。首先,利用局部更宽残差块结构,设计了两级残差特征提取模块;然后,以该模块为基础,使用特征图循环传递的方式来构造深层特征提取网络,这可以使得多个网络层共享参数,提高了网络的效率;最后,改进了以往惯常使用的上采样方法,为了弥补分辨率扩张带来高频信息的损失,采用多尺度联合学习的机制构建上采样模块。实验结果表明,与相同类型网络相比,本文方法在维持网络规模较低时,获得了优秀的性能指标和视觉效果。

     

    Abstract: Deep learning-based methods for image super-resolution reconstruction often have a large model scale and low efficiency. Therefore, we propose an image super-resolution reconstruction network that can balance performance and network size well. We design a dual-level residual feature extraction module by using the local wider residual block structure. Using this module as a basis, we construct a deep feature extraction network by the feature map cyclic transfer method, which allows multiple network layers to share parameters and improve the efficiency of the network. In addition, we construct an upsampling module using a multiscale joint learning mechanism, which improves the commonly used upsampling method to compensate for the loss of high-frequency information caused by resolution expansion. Experimental results show that compared with the same type of network, the proposed network achieves excellent performance indicators and visual effects while maintaining a small network size.

     

/

返回文章
返回