This project focuses on developing a denoising fully CNN Auto-encoder based architecture, on the CIFAR-10 dataset. Different architectures are examined and finally two of them are tested using a Visual Transformer fine tuned on the CIFAR-10 dataset. More details about this work can be found on the Denoise.pdf file.
In the source/main.py file one can train the model based on the specifications provided by the tests_fully.csv. The noise is added using the source/noisy.py file. The architectures developed can be found on source/model_b.py and source/model_c.py. Model-C is the best model developed so far, but it can be made even better. All progress steps are logged on the Notes_on_training.odt file, so anyone can have an overview over the whole process. The best model checkpoints (test ids: 35, 37 and 38 as refered in tests_fully.csv) can be found on the archive folder. All test results, that is Loss and PSNR plots can be found there as well, but only for the three best models there exist the .pt pytorch model checkpoints.
In the figures below it can be seen the output of the model with the final image distribution per channel.
For the training two different types of noises were used and their combinations. More of the noise types can be found on the source/noisy.py file.
Processor: Intel® CoreTM i7-8550U, 4 GHz (turbo mode)
GPU: Nvidia MX150, 2 GB
RAM: 16 GB
Operating System: Ubuntu 23.10
e-mail: chrispsyc@yahoo.com