Convolutional Neural Networks for Rate-Distortion (CNN-RD).
Coding Framework complemented with the two following Convolutional Neural Networks (CNNs).
- CNN-CR for down-sampling before image coding.
- CNN-SR for up-sampling afer image decoding.
This framework allows to train both CNNs with a loss function that minimizes both distortion (with MSE) and rate. The former is achieved by estimating the Discrete Cosine Transform coefficients that JPEG would quantize to zero.
This project was developed at Instituto de Telecomunicações (IT) and Instituto Superior Técnico in a Master Thesis context.
Below are the instructions to run the provided framework.
Run train.py without any arguments. Datasets, Hyper-parameters and settings are all hardcoded and defined at the beggining of the script. All these settings are commented to help change them if necessary.
During training, the obtained models are evaluated every epoch.
For indepedent inferece (i.e. without running the training script) run eval.py without any arguments. Datasets and settings are all hardcoded and defined at the beggining of the script. All these settings are commented to help change them if necessary.
For any question or problem, please contact [email protected] or open an issue.