| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 107 === Recent deep learningbased approaches have shown significant improvements in image completion. However, the existing methods often create distorted structures or blurry textures inconsistent with surrounding areas. This is mainly due to the ineffectiveness of convolutional neural networks in copying information from distant spatial locations. Therefore, the contour cues in the image are quite important, and we can more definitely confirm the boundary between objects by this information. In this thesis, we propose a two-stage architecture for image completion, which is the edge completion network and the coarsetofine image completion network. Edge completion network generates edges in missing regions, we use hinge loss for training to determine whether the input is real or fake, it also makes the completed edges more realistic. Then, image completion network generates an image from low dimension to high dimension, the completed edges are as a condition fed into both coarse network and refine network that makes the boundary of the missing parts more reasonable, meanwhile increase larger receptive field. Experiment results show that our method can generate better quality images than the state-of-art approaches in both quantitatively and qualitatively.
|
|---|
