Dual-Flow Feature Coupling GAN for Image Inpainting

Image inpainting is an significant task in the field of computer vision, which can generate exquisite images from incomplete input images. However, insufficient consideration of the relationship between global and local features during the inpainting process can result in poor consistency or even blurring on the generated image contents. In view of this, a dual-flow feature coupling GAN (DFC-GAN) is proposed, which consists of two encoding and decoding subnetworks, namely the main stream and Transformer stream. The main stream is designed to adopt the fully convolutional network for better capturing features. The Transformer stream is presented to use stacked fast Fourier Transformers (FFTr) to expand the receptive fields effectively. Besides, a space and channel aggregation (SCA) module is introduced to fuse local features extracted by the main stream and global features captured by the Transformer stream to avoid potential domain conflicts. A global feature transfer (GFT) module is designed to couple global features generated by the Transformer stream and features from the main stream layer by layer, which can guide the generation of main stream features. Extensive experiments on two public datasets, including Paris StreetView and CelebA-HQ, demonstrate that the proposed method outperforms the state-of-the-art methods both quantitatively and qualitatively.