Modified an inception model and trained it on 10K images to colorize grayscale pictures. As for color space, CIELAB is used to separate lightness as input and a* and b* as outputs of the convolutional neural network. (Color masks below are a* and b* together with L=50 to illustrate raw output)
| Input Grayscale Image | Colorized Output Image | Tuned Saturation | Raw Output Color Mask |
|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
|
 |
Comparison with Zhang Colorization
Please note that as mentioned in Zhang's paper, their model was trained on over a million color images; however, I only had 10,000 pictures.
| Input Grayscale Image | My Result | Tuned Saturation | Zhang et al. |
|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
The best results were observed when I used the Inception network's layers with weights of Imagenet, and I replaced the last 19 layers by a small CNN, including Conv2d, BatchNormalization, and Conv2DTranspose layers to recreate a picture with the same size as input. For 100 epochs, I fixed the weights, but after showing no more improvement, I made all the parameters trainable. The results above are from about 100 epochs after that change.