Image-Colorization

Running the Model

1. Download Places365 to the Dataset directory.
2. Extract places365_train_standard.txt from places365_train_standard.zip to the Dataset directory.
3. Start training the model by running all the blocks in UNet_PatchGan_v4.ipnyb.
4. To evaluate the model with new pictures, modify ckpt_path and img_path in the third and last blocks of Evaluate.ipnyb respectively.

To change the root directory of datasets, modify dataset.py on Line 8 and 34. Be sure to include places365_train_standard.txt in the new directory.

Things to Note

1. The model is mainly based on PatchGan. A classifier and Convolution Block Attention Module (CBAM) are incorporated to improve the model's performance.
2. For this model, pictures are represented in LAB color space. There are two main reasons why we utilize LAB instead of RGB: Given that the "L" channel can be used as input, the model only needs to generate and concatenate values in the "A" and "B" channels; on the other hand, it will be easier for the model to colorize images of varying size, since the rescaled "A" and "B" channels can be concatenated using the original "L" channel with no distortion in pictures.

Project Paper