This project is an implementation practice of paper Deep Painterly Harmonization. This network allows you to copy and paste some part of image element from other image to your target image and smooth the artifacts, such as:
Only the first step is implemented owing the time limitation. I will finish the second part after current research work.
Owing to the limitation of memory and GPU, I cannot reconstruct a large image. Please check the code and select the suitable size of output.
- python 3.x
- Tensorflow
- Keras
Before reconstructing the image, you should send the materials requied by the training process. Samples of data organization can be finded in input/painting/xxx/
- Pasted input, which is shown above.
- bw_mask, which tells the network which part is the target and which part is the resource.
- backgound, the original image
To run the training, directly run the code paint-train.py. You would like to check the code and switch the path of your dataset. Assuming that you have installed all the required libs and your Keras has the pretrained VGG-19 model.
python paint-train.py
These results can be checked in "input/painting". The left image is the original background, the mid one is the copy-and-paste input, and the right one is the reconstructed image. Considering that the results are generated through first step, the results will not be satisfying enough. Also the size of results will be small owing to the limitation of memory and GPU. You can modify the size of output in code and I will try to implement the second part later.
