In this Reposrity,I used Autoencoder to test 22 differents models to find the best architechture that suite image denoising problem. I tries Convolution Neural Network with different parametres (kernel size, filter) as well as Fully connected layer with different Hidden unites.
The notebook contains all the code without the logs because I have trouble with timout in colab (Runing the notebook may take more then 6 hours that is why you find only the logs of the first 13 models ) . SO I highly recommand you if you have GPU in your local machine and you want to see all the results of each model, run all the cells .
Make sure you have installed all the libraries :
We used Two different datasets from kaggle, Smartphone Image Denoising Dataset (SIDD) and
Super Resolution Benchmarks .
The first one has the noised already but for the second one(Super Resolution Benchmarks) I have to build function that added noises to each one of the entire dataset.
We merge the two after adding noises to the second one and then we implement tf.data.Dataset.from_tensor_slices
to load the data to the models .
if you prefer to work on kaggle you can check the notebook here
Tensorflow is the framework that we used in the notebook, and we manage to build 22 different models to see what architechture will suite image denoising problem .
for evaluation we use Mean square error (MSE) and Mean absolute error (MAE) TODO: metrics results TODO : bar chart