Face Swapping Neural Networks Based on Improved Autoencoders

Traditional face swapping method is manually synthesized through image editing software such that the synthetic results are not only unnatural but also time consuming. This paper proposes a face swapping system mainly based on the improved autoencoders such that the swapped faces can be achieved more natural results and faster speed. In the proposed face swapping system, MTCNN and face recognition are first used to detect and collect the faces of characters A and B from two different video sources. With the collected two facial data sets for characters A and B, we can employ the proposed shared-encoder to extract their facial information and restore their faces by the proposed decoders A and B, respectively. The trained neural networks can swap the faces of character A to the faces of character B if we use shared-encoder with decoder B. Finally, by image fusing process, the swapped faces are pasted to the original video to finish the face swap procedure. The experimental results are exhibited to show the effectiveness of the proposed face swapping neural network system.