Automatic Gaze Correction based on Deep Learning and Image Warping
Masataka Seo, Takahiro Yamamoto, Yen‐Wei Chen, Toshihiro Kitajima · 2019
When people take a selfie photo or talk through a video chat system, they tend to look at the screen. Since the position of the camera is usually different from that of the screen, the eyes of the person in the photograph or transmitted live video seem to make contact with something other than the viewer. This paper proposes a method to correct the gaze by synthesizing the image texture. The image synthesis uses feature points around the eyes as landmarks. A common convolutional neural network and long-term recurrent convolution network are used to extract or detect these feature points in a still image and video image, respectively. The deep learning-based feature point detections are very accurate compared to conventional methods. These feature points are used as landmarks, and the internal texture of these feature points are synthesized with the prepared template image in advance. Through these procedures, an automatic and natural gaze correction is realized. This method realized a natural gaze correction with little blur, even for a video image.