Generative Image Inpainting for Fine Details

Abstract

Since the rapid development of deep learning, image inpainting techniques have also improved significantly. Although these techniques have been able to reconstruct semantically coherent and visually plausible masked regions compared to traditional techniques, the results of these works are commonly blurry due to lack fine details. This paper proposes a novel model including an image completion network and an edge matching module. The image completion network is a Generative Adversarial Network framework added skip-connection for contextual feature fusion, and the edge matching network facilitates the image inpainting network by constraining the edge of results. We evaluate our model on the publicly available datasets CelebA and Places2. Results show that our method outperforms current representative technique quantitatively and qualitatively.