Laplacian pyramid adversarial network for face completion

摘要: Abstract Recently, generative adversarial networks (GANs) have demonstrated high-quality reconstruction in face completion. There is still much room for improvement over the conventional GAN models that do not explicitly address texture details problem. In this paper, we propose a Laplacian-pyramid-based framework This can produce more realistic results (1) by deriving precise content information of missing regions coarse-to-fine fashion and (2) propagating high-frequency from surrounding area via modified residual learning model. Specifically, regions, design convolutional network predict under different resolutions; takes advantage multiscale features shared low levels extracted middle layers next finer level. For details, construct new to eliminate color discrepancies between progressively. Furthermore, multiloss function proposed supervise process. To optimize model, train entire model with deep supervision using joint loss, which ensures generated image as original. Extensive experiments on benchmark datasets show exhibits superior performance state-of-the-art methods terms predictive accuracy, both quantitatively qualitatively.