Image Inpainting Based on Dual Dynamic Predictive Filtering and Fast Fourier Convolution

Despite outstanding performances of the existing image inpainting methods in restoring corrupted images, they often generate images with artifacts or missing details. Predictive filtering is a widely used image restoration technique which can preserve local structures and avoid artifacts, but it cannot fill large missing areas. Convolutional neural networks are suitable for large missing areas, but it is difficult to restore fine details. Therefore, we treat image inpainting as a filtering task, and propose a novel image inpainting method consisting of two networks, namely the Attention-Enhanced Kernel Predictive Network (AE-KPNet) and the Deep Generative Network (DGNet). The two parallel networks are linked, and DGNet provides features to AE-KPNet, while AE-KPNet trains dual dynamic predictive filtering kernels for DGNet. To address the issue of large missing areas, we introduce fast Fourier convolution into DGNet, and the operation in the frequency domain expands the spatial receptive field. We evaluate our model on three public datasets, CelebA, Places2, and Paris StreetView. Through quantitative and qualitative experiments, the results show that our method achieves the performance of state-of-the-art methods.