Sample-based image completion using structure synthesis

• It is a difficult topic that filling in large scale missing area in images naturally.
• Using of foregrounds of source and sample images can get rid of useless information.
• Global contour matching using BBM vector is devised to determine transform function.
• Structure is reconstructed according to optimal sample patch and local consistency.
• Weighted exemplar-based image synthesis combines source and sample images’ effects.

Image completion technique is widely used in image processing applications such as textural recovery, object removal, image edit, etc. When filling in the missing areas of an image, it is often a challenge to keep local consistency of image structures while avoiding ambiguity and visual artifacts. To tackle with this problem, we propose a robust sample-based image completion scheme which is a cascade of two major procedures. First, we extract structural information from both source and sample images and then apply boundary band map (BBM) descriptor to perform template matching under contour consistency constraint and reconstruct the damaged structures. Second, a weighted exemplar-based image synthesis algorithm is further devised taking the previous structural information and matching results into account. Extensive experiments and comparative study show the reliability and superiority of our image completion algorithm.