WFIL-NET: image inpainting based on wavelet downsampling and frequency integrated learning module

The purpose of image inpainting is to restore and fill missing areas, and how to restore delicate and reasonable missing content has always been one key issue. In the past decade, remarkable achievements have been made in image inpainting based on deep learning. However, when faced with large and irregular missing areas, there are still some problems such as semantic inconsistency, blurred edges and artifacts in the inpainted images. To address these problems, this paper proposes a novel image inpainting algorithm WFIL-NET which is based on wavelet downsampling and frequency integrated learning module. The WFIL-NET adopts the generative adversarial network (GAN) structure, where the Encoder–Decoder network is used in the generator part. To retain rich information while reducing the image resolution, we propose to use wavelet downsampling module in the encoder part to enhance the capacity of subsequent operations to learn representative features. Moreover, the wavelet transform extracts image features at different frequency levels: low-frequency information encapsulates the primary content and structure, whereas high-frequency information captures details and texture. The proposed frequency integrated learning module employs the attention mechanism to allocate appropriate weights to high and low frequency information, effectively integrating them to ensure a more coherent structure and semantic consistency in the inpainted image. Experimental results on the CelebA-HQ and Places2 datasets demonstrate that the proposed method effectively fills large and irregular missing areas, significantly enhances the visual quality of inpainted images, and mitigates edge blurring and artifacts.