Image super resolution

Introduction

this project is a reproduction of the famous paper Learning a Deep Convolutional Network for Image Super-Resolution which was publish in 2014 using CNN to do super resolution work

• paper pdf-download

Result

low-resolution	high-resolution

Related work

• A CSDN blog that describes details about what is super resolution
• An excellent github, this repository was based on it, it offers me great help.
• great job!!! strongly recommend

Requirements

• pytorch

• pip install -r requirements.txt
  

Use

Step1: Prepare the dataset

download the standard dataset The 91-image(train set), Set5(test set) dataset converted to HDF5 can be downloaded from the links below.

Download any one of 91-image and Set5 in the same Scale and then move them under ./datasets as ./datasets/91-image_x2.h5 and ./datasets/Set5_x2.h5

Step2: Train the model

• easy run

  python train.py
  

• about arguments

  • train-file eval-file 2/3/4, different datasets to choose
  • batch-size
  • num-workers
  • lr learning rate
  • epoch
  • f frequency to test the model
  • model-dir where the model was saved

  python train.py --train-file 4 --eval-file 4 --batch-size 64 --lr 1e-5 --num-workers 8 --epoch 500 --f 10 
  

  All model will be saved under ./model and the best model is ./model/best.pth

Step3: Download the pretrained model(**If you have trained your model, skip this part)

• trained by x2
• trained by x4

Move the model weights under ./model as ./model/best.pth or ./model/b.pth

Step4: Do super resolution

python use.py --weights-file ./model/best.pth --image x/xx/xxx.jpg

By default, the weights file is ./model/best.pth, if you want to use b.pth please replace it. A picture will be created named as xxx_srcnn.jpg

Conclusion

This is just a CNN try for me, to familiar with basic steps of machine learning, almost all code comes from here,I just follow his step and reapperance his job. The final result couldn't satisfy me, it actually should be called little super resolution haha. However, there's no doubt that idea of SRCNN was novelty in that period. If you want to improve the neural network, build a deeper model and try Resnet model.

If you are interested in it, you could see this great job and i strongly recommend it.

some errors you may occur

• OSError: Unable to open file (file locking disabled on this file system (use HDF5_USE_FILE_LOCKING

  Solution:

  #linux
  nano ~/.bashrc
  

  Add export HDF5_USE_FILE_LOCKING='FALSE' in a line, use Ctrl-X to exit

  source ~/.bashrc
  

• OSError: [WinError 1455] 页面文件太小，无法完成操作。 Error loading "C:\ProgramData\Anaconda3\envs\study_and_test\lib\site-packages\torch\lib\caffe2_detectron_ops_gpu.dll" or one of its dependencies. Solution:

  python train.py --num-workers 0