High-Resolution Gaze-Corrected Image Generation based on Combined Conditional GAN and Residual Dense Network

In recent years, the ease of taking selfie photos has improved due to widespread use of smartphones and social networking services (SNSs). In a typical smartphone, the camera layout and display are different, rendering the gaze often not front-facing. Although various gaze-correction methods have been proposed for this problem, many of them cannot generate sufficiently natural images. Furthermore, no improvement can typically be seen in the unnatural images generated with these methods. In this paper, we propose a gaze-correction method using a deep generative model. This model can determine the naturalness of the resulting images and learn to provide natural results. In addition, recent smartphones have the ability to take large-sized photos. However, it is difficult for this model to generate large-sized images. In this paper, we use super-resolution techniques to generate images that have a larger size and higher resolution than those generated by conventional methods. In our experiments, to reduce the size of our neural network, we input a small-sized image and convert it back to its original size after performing gaze-correction.