Structure Guided Image Inpainting using Edge Prediction

Kamyar Nazeri1
Eric Ng1
Tony Joseph2
Faisal Z. Qureshi2
Mehran Ebrahimi1
Imaging Lab1 Visual Computing Lab2
Faculty of Science1,2
University of Ontario Institute of Technology
2000 Simcoe St. N., Oshawa ON L1G 0C5

Abstract

We propose a two-stage model that separates the inpainting problem into structure prediction and image completion. Similar to sketch art, our model first predicts the image structure of the missing region in the form of edge maps. Predicted edge maps are passed to the second stage to guide the inpainting process. We evaluate our model end- to-end over publicly available datasets CelebA, CelebHQ, Places2, and Paris StreetView on images up to a resolution of 512 x 512. We demonstrate that this approach outper- forms current state-of-the-art techniques quantitatively and qualitatively.

Proposed Generative Adversarial Network for structure guided image completion using edge maps. Incomplete grayscale image and edge map, and mask are the inputs of G1 to predict the full edge map. Predicted edge map and incomplete color image are passed to G2 to perform the inpainting task.
Proposed Generative Adversarial Network for structure guided image completion using edge maps. Incomplete grayscale image and edge map, and mask are the inputs of G1 to predict the full edge map. Predicted edge map and incomplete color image are passed to G2 to perform the inpainting task.

Some results

The proposed network is able to use color and edge information from different sources to create unique images.

Edge-map (3rd column) generated using the left-half of (1st column) (shown in black) and right-half of (2nd column) (shown in red). Input is (1st column) with the right-half removed, producing the output (last column).
Edge-map (3rd column) generated using the left-half of (1st column) (shown in black) and right-half of (2nd column) (shown in red). Input is (1st column) with the right-half removed, producing the output (last column).

The proposed method can also be used for photo editing as seen below.

Examples of object removal and image editing us- ing our EdgeConnect model. (Left) Original image. (Cen- ter) Unwanted object removed with optional edge informa- tion to guide inpainting. (Right) Generated image.
Examples of object removal and image editing us- ing our EdgeConnect model. (Left) Original image. (Cen- ter) Unwanted object removed with optional edge informa- tion to guide inpainting. (Right) Generated image.

Supplementary Material

More information is available here.

Publication

For technical details please look at the following publications