Enhanced Deep Residual Networks for Single Image Super-Resolution
- Abstract:
- Use Residual Networks but optimize it by analyzing and removing unnecessary modules
- Do some modification on training multi-scale training
- PSNR and SSIM
- Related work:
- Interpolation base on sample theory:
low resolution image is down sample of the high resolution image- Adopt nature image statistic to problem (?)
- Neighbor embedding (machine learning method)
- Sparse coding (machine learning method)
- Deep Neural Networks:
- SRCNN 香港中文大学,汤晓鸥:
3 convolutional layers:
feature extraction --> non-linear mapping --> reconstruct
- Residual Network for SR:
To get deeper network and superior performance
- Use Recursive Convolutional Network:
《Deeply-Recursive Convolutional Network for Image Super-Resolution》
alleviate the burden of carrying identity information in the super-resolution network???.
- Encode and Decode structure with skip connections:
they said it provide fast and improved convergence
- Training strategy:
- Many method do the upsample operation via bicubic Interpolation before the image feed to the Network
- In VDSR, they train a single Network using multi-scale images 2X 4X 8X so the network can deal with multi-scale problem.
- Geometric Self-ensemble, do some transform operation to the lr images as a set and then feed them to the network(个人理解是生成多个针对于不同角度变换lr的学习模型,然后将这些模型给一些权重,集成成一个更具有鲁棒性的学习器)
- Loss function: MSE or L2 loss, SSIM
- Propose method:
remove BN: they get rid of range flexibility from
networks by normalizing the features ??? - 总体来说这篇文章是在resnet网络的基础上对网络的多余部分进行了精简,例如删掉BN层等操作,并且使用了以前提到的multi-scale训练策略来进行实验。