SRCNN, Image Super-Resolution Using Deep Convolutional Networks in Tensorflow.
My implementation of Image Super-Resolution Using Deep Convolutional Networks in Tensorflow.
The algorithm is not mine, its an implementation of the SRCNN introduced by Dong et al
I have supplied a few functions to visulise the weights of the model (kernels) and the feature maps.
Project Structure:
- model - you can find the preon trained model, model is trained on a single colour channel, and hence expects this as as in input
- srcnn.py - here I initialise the network, and test it on image supplied in "image" directory
- outputs - self explanatory, you can check the results obtained in here
The project is using Python 3.5.
Network
-
Convolitional Neural Network
-
Filter Sizes: 9 - 1 - 5, as in the original paper
-
Learning rate is 10^−4 for the first two layers, and 10^−5 for the last layer, according to the authors' benchmark
List of Dependencies:
-
MatPlotLib - for visualisation
-
NumPy - for everything, its NumPy
-
Tensorflow - network built
-
Scipy - for image processing
-
Scikit-image - computing PSNR
NOT JUPYTER NOTEBOOK
Run from terminal, PyCharm, or from anything else that makes you happy:)
Everything else you need to know is in the comments.
Star, fork, do your thing.