Super-resolution imaging using DNNs

Installing dependencies

Anaconda3 is recommended as the default python enviroment.
Run pip from the terminal/command prompt on the requirement file to install the needed dependencies. For cpu based tensorflow :
```
foo@foobar:~$ pip install -r requirements-cpu.txt
```
For Tensorflow GPU support, first need to install CUDA 9.0 and CuDNN 7.1. See https://www.tensorflow.org/install/gpu for instructions. Then run :
```
foo@foobar:~$ pip install -r requirements-gpu.txt
```

First, need to create hdf5 databses which are very space efficient and quick to read from disk
Use makeHDF5.py to create these files. You must provide the path to the config_db.toml file which contains the settings for your database
```
foo@foobar:~$ python makeHDF5.py ./config_db.toml
```
Some key settings inside the config_db.toml file :
- image_sets ~ a list of paths to folders containing the images to include in the database
- kfold ~ the number of cross folds. If kfold is set to 5, the script will generate 5 hdf5 files each with randomly distributed patches from different image files. Patches from any image are found only in one hdf5 file.
- num_patches ~ the number of patches to take from each image
- original_patches ~ the crop size for each patch. Can list several sizes, but they must be smaller than original image size. All patches are square.
- resized_patches ~ the size to which each of the original patches are resized to. Can list several sizes, but only sizes where original size is greater will be taken into account.
- normalize ~ if set to true, all image patches will be scaled to values between 0.0 and 1.0.

Models are defined in modelLib.py
Some configuration of the model like input size and output size etc. can be parameterized and set from config_train.toml for fast experimentation

Once you have hdf5 files and models defined, use train.py to train your models
```
foo@foobar:~$ python train.py ./config_train.toml
```
Some key settings inside the `config.toml are :
- x ~ this is the name of the database entry in the hdf5 files to use as input (e.g. org_64_rsz_32 for patches of size 64 x 64 from the original image which were resized to 32 x 32)
- y ~ this is the name of the database entry in the hdf5 files to use as the target for the model (e.g. org_64 for patches of size 64 x 64)
- train ~ list of paths to hdf5 files which will be used as training set. If more than one path provided, cross validation will be performed
- val ~ list of paths to hdf5 files which will override the default test set during cross validation
- modelArch ~ the name of the model as defined in modelLib.py
- modelDir and modelName ~ path at which to save model weights and other results