Portrait Shadow Manipulation

Remove foreign shadows and soften facial shadow in a portrait photo based on GridNet.

Overview

I implemented a Neural Network to removed the foreign shadow on human face in a portrait image based on the paper Portrait Shadow Manipulation.

In brief, the method proposed in the paper relies on a pair of neural networks—one to remove foreign shadows cast by external objects, and another to soften facial shadows cast by the features of the subject and to add a synthetic fill light to improve the lighting ratio. To train the first network, they constructed a dataset of real-world portraits wherein synthetic foreign shadows are rendered onto the face.


Network Architecture

For foreign shadow removal model, a GridNet architecture with modifications was employed. GridNet can learn both high-level and low-level features with its multi-resolution conv-deconv grid architecture which can be seen as an extension of the U-Net.

Besides, the model is supervised with a weighted combination of pixel-space L1 loss (\(L_{pix}\)) and a perceptual feature space loss (\(L_{feat}\)). The perceptual loss is computed by processing the images through a pre-trained VGG-19 network and computing the L1 difference between extracted features in selected layers.

Data Acquisition

To synthesize images with foreign shadows, we model images as a linear blend between a “lit” image \(I_l\) and a “shadowed” image \(I_s\), according to some shadow mask \(M\):

\[I=I_l\circ (1-M)+I_s\circ M\]



Since style-GAN generated images are lifelike and quite similar to human protrait images, I used it to generate 50,000 different identities and took the method above to synthesize the foreign shadow dataset.


Results

The testing data was sythesized in the same way with training data, but with different identities and different shadow shapes. The final model can perform very well on the testing data, but there will be obvious masks on the real-world shadow portaits.

Results on the Sythesized Testing data.
Results on the real-world data.