计算机科学与探索 ›› 2023, Vol. 17 ›› Issue (4): 879-891.DOI: 10.3778/j.issn.1673-9418.2107136

• 图形·图像 • 上一篇    下一篇

多尺度自适应上采样的图像超分辨率重建算法

吕佳,许鹏程   

  1. 1. 重庆师范大学 计算机与信息科学学院,重庆 401331
    2. 重庆师范大学 重庆市数字农业服务工程技术研究中心,重庆 401331
  • 出版日期:2023-04-01 发布日期:2023-04-01

Image Super-resolution Reconstruction Algorithm Based on Multi-scale Adaptive Upsampling

LYU Jia, XU Pengcheng   

  1. 1. College of Computer and Information Sciences, Chongqing Normal University, Chongqing 401331, China
    2. Chongqing Center of Engineering Technology Research on Digital Agriculture Service, Chongqing Normal University, Chongqing 401331, China
  • Online:2023-04-01 Published:2023-04-01

摘要: 针对现有基于卷积神经网络的图像超分辨率重建网络中存在的训练难以收敛、无法适配多个放大系数和无法进行非整数放大系数上采样的问题,提出了一种基于多尺度和分而治之思想的自适应上采样图像超分辨率重建算法。该算法通过改进后的多尺度通道注意力特征提取模块对低分辨率图像进行多尺度特征提取以生成不同尺度下的特征图,再将其输入瓶颈层实现全局特征融合,使用基于分而治之的自适应上采样模块获得超分辨率图像,从而解决了不同放大系数的适配问题和非整数放大系数的上采样问题。在对比实验中,该算法在不使用任何初始化方法时仍然具有良好的收敛性。在整数放大系数下,该算法的图像重建性能超过当前主流的超分辨率网络,PSNR和SSIM性能相比MRFN分别提升了0.34 dB和0.039 1。在非整数放大系数下,其PSNR性能相比双三次插值方法平均提升1.24 dB,且不需要对每一个放大系数都进行训练。

关键词: 图像处理, 超分辨率重建, 卷积神经网络, 多尺度特征提取, 分而治之

Abstract: In view of the problems in the existing convolutional neural network-based super-resolution network, such as difficult convergence of training, unable to adapt to multiple upsampling coefficients and unable to sample non-integer upsampling coefficients, an image super-resolution reconstruction algorithm of adaptive upsampling module based on multi-scale and the idea of divide-and-conquer is proposed in this paper. In the algorithm, the improved multi-scale channel attention feature extraction module is used to extract multi-scale features of low-resolution images to generate feature maps at different scales, and then the global feature fusion is realized by inputting them into the bottleneck layer. The super-resolution images are obtained by using the adaptive upsampling module based on divide-and-conquer to solve the adaptation problem of different upsampling coefficients and the upsampling problem of non-integer upsampling coefficients. In the contrast experiment, the proposed algorithm still has good convergence without any initialization methods. Under the integer magnification factor, the image reconstruction performance of the proposed algorithm exceeds the current mainstream super-resolution network, and the PSNR and SSIM performance are improved by 0.34 dB and 0.0391 respectively compared with MRFN. Under the non-integer magnification factor, the average PSNR performance is improved by 1.24 dB compared with the bicubic interpolation method, and there is not necessary to train each magnification factor.

Key words: image processing, super-resolution reconstruction, convolutional neural network, multi-scale feature extraction, divide-and-conquer