Machine Learning for APS360
This project looks to create an CNN based autoencoder model using several techniques such as residual connections and instance awareness to colorize black and white images
A landscape dataset from kaggle is used for the same. We pre-process images to 128 x 128 with padding or crop, and use the LAB color space. The pre-processed image is saved as a numpy array which is used to train the model using a custom DataLoader.
We use a UNET style architecture for the autoencoder with a CSPDarknet52 object detection backbone (pretrained) embedding fused to the model at the bottleneck.
PSNR Loss is predominantly used for training.
Several helper functions such as quick_qualitative, to_rgb and various others can be found in the notebook too.
We achieve a PSNR of approximately 20.
Use APS360_Final for in Colab to replicate results!
Hyperparamater Testing Notebook https://drive.google.com/file/d/1w747clyX6h948jA74_4gKZK5-R6fBJBc/view?usp=sharing
Test results: https://colab.research.google.com/drive/10HdZ-0Ep1f0jkc0knoUAnrLq9j9yB0rH?authuser=8#scrollTo=N8xyG11z-Tih