期刊文献+

多尺度残差和二阶退化的图像超分辨率重建 被引量:1

Image Super-resolution Reconstruction with Multi-scale Residuals and Second-order Degradation
下载PDF
导出
摘要 针对目前图像超分辨率重建算法中因退化过程过于单一所导致的网络性能下降和模型泛化能力差等问题,本文提出了多尺度残差和二阶退化的图像超分辨率重建算法。该算法首先设计了二阶退化模型,在每一阶退化过程中加入随机的下采样、模糊、噪声和压缩操作以保证退化模型的复杂性和易用性。其次提出了多尺度感受野残差密集块,利用多分支结构和空洞卷积来增强网络的特征提取能力。最后改进了上采样方式,交替使用双线性插值和亚像素卷积上采样算法,以平衡算法性能和时间复杂度。实验结果表明,该算法在三个基准数据集上的自然图像质量评估指标平均下降了1.15,且重建图像视觉观感更好,纹理细节、亮度和饱和度更加准确。 To address the problems of degraded network performance and poor model generalization caused by too single degradation process in current image super-resolution reconstruction algorithms,this paper proposed an image super-resolution reconstruction algorithm with multi-scale residuals and second-order degradation.The algorithm firstly designed a second-order degradation model,and added random downsampling,blurring,noise and compression operations in each order of degradation to ensure the complexity and ease of use of the degradation model.Secondly,multi-scale receptive field residual dense block is proposed,which utilized multi-branch structure and null convolution to enhance the feature extraction capability of the network.Finally,the upsampling method was improved by alternating the bilinear interpolation and sub-pixel convolutional upsampling algorithms to balance the algorithm performance and time complexity.Experimental results show that the algorithm achieves an average reduction of 1.15 in natural image quality evaluator metrics on three benchmark datasets and better visual perception of the reconstructed images with more accurate texture details,brightness and saturation.
作者 赵杨坤 章义来 肖佳涛 ZHAO Yangkun;ZHANG Yilai;XIAO Jiatao(School of Information Engineering,Jingdezhen Ceramic University,Jingdezhen,China,333403)
出处 《福建电脑》 2023年第6期6-12,共7页 Journal of Fujian Computer
基金 潮州市陶瓷产业人才振兴计划(No.2021YJ03)资助。
关键词 超分辨率 残差神经网络 二阶退化 图像处理 Super-resolution Residual Neural Network Second-order Degradation Image Processing
  • 引文网络
  • 相关文献

参考文献1

二级参考文献4

共引文献1

同被引文献4

引证文献1

二级引证文献1

;
使用帮助 返回顶部