基于多尺度方形Densenet的图像超分辨算法
|
摘要
近年来,基于CNN的单张图像超分辨(SISR)算法取得了巨大的进步。目前基于CNN的SISR算法都是以单尺度CNN网络为主,但是这种网络结构难以有效利用图像的多尺度信息。针对这一问题,本文提出了一种基于多尺度方形Densenet(MSR-Densenet)的SISR算法。首先,设计了一种多尺度方形(MSR)网络,使得CNN网络能够有效利用图像多尺度信息;其次,将Densenet的稠密连接思想引入MSR网络,提升算法的收敛性;最后,定义多尺度损失函数,使得算法在测试时能够对网络自适应剪枝以降低计算复杂度。实验结果表明,MSRD算法明显优于传统超分辨方法(PSNR和SSIM分别提升0.8dB和0.1),而且MSRD能够在保证超分辨性能的前提下,通过自适应剪枝有效降低计算量。
In recent years, CNN-based single image super-resolution(SISR) algorithms have made great progress.At present, the CNNR-based SISR algorithms are mainly based on single-scale CNN networks, but such network structures are difficult to effectively utilize multi-scale information of images. To solve this problem, this paper proposes a SISR algorithm based on multi-scale rectangle Densenet(MSR-Densenet). Firstly, a multi-scale Rectangle(MSR) network is designed to make the CNN network effectively use image multi-scale information.Secondly, Densenet's dense connection idea is introduced into the MSR network to improve the convergence of the algorithm. Finally, the multi-scale loss function is defined so that the algorithm can adaptively prune the network during testing to reduce computational complexity. The experimental results show that the MSRD algorithm is significantly better than the traditional super-resolution method(PSNR and SSIM respectively improve 0.8 dB and 0.05), and MSRD can effectively reduce the amount of computation by adaptive pruning under the premise of ensuring super-resolution performance.
In recent years, CNN-based single image super-resolution(SISR) algorithms have made great progress.At present, the CNNR-based SISR algorithms are mainly based on single-scale CNN networks, but such network structures are difficult to effectively utilize multi-scale information of images. To solve this problem, this paper proposes a SISR algorithm based on multi-scale rectangle Densenet(MSR-Densenet). Firstly, a multi-scale Rectangle(MSR) network is designed to make the CNN network effectively use image multi-scale information.Secondly, Densenet's dense connection idea is introduced into the MSR network to improve the convergence of the algorithm. Finally, the multi-scale loss function is defined so that the algorithm can adaptively prune the network during testing to reduce computational complexity. The experimental results show that the MSRD algorithm is significantly better than the traditional super-resolution method(PSNR and SSIM respectively improve 0.8 dB and 0.05), and MSRD can effectively reduce the amount of computation by adaptive pruning under the premise of ensuring super-resolution performance.
引文
[1]Harris J L.Diffraction and Resolving Power[J].Journal of the Optical Society of America(1917-1983),1964,54(7):931-933.
[2]苏衡,周杰,张志浩.超分辨率图像重建方法综述[J].自动化学报,2013,39(8).
[3]罗山.基于多尺度几何分析的图像增强方法综述[J].山西电子技术,2015(2):94-96.
[4]Dong C,Loy C C,He K,et al.Learning a Deep Convolutional Network for Image Super-Resolution[C]//European Conference on Computer Vision.Springer,Cham,2014.
[5]Yan Y,Zhang L,Li J,et al.Accurate Spectral Superresolution from Single RGB Image Using Multi-scale CNN[J].2018.
[6]Molchanov P,Tyree S,Karras T,et al.Pruning Convolutional Neural Networks for Resource Efficient Inference[J].2016.
[2]苏衡,周杰,张志浩.超分辨率图像重建方法综述[J].自动化学报,2013,39(8).
[3]罗山.基于多尺度几何分析的图像增强方法综述[J].山西电子技术,2015(2):94-96.
[4]Dong C,Loy C C,He K,et al.Learning a Deep Convolutional Network for Image Super-Resolution[C]//European Conference on Computer Vision.Springer,Cham,2014.
[5]Yan Y,Zhang L,Li J,et al.Accurate Spectral Superresolution from Single RGB Image Using Multi-scale CNN[J].2018.
[6]Molchanov P,Tyree S,Karras T,et al.Pruning Convolutional Neural Networks for Resource Efficient Inference[J].2016.