An Dual-Branch Network Based on Wavelet Decomposition for Image Inpainting

In this paper, we propose the Wavelet-Guided Dual-Branch Network (WGDN), an innovative deep learning framework designed for image inpainting. WGDN uniquely leverages wavelet decomposition to address the challenge of reconstructing large missing regions by analyzing both low-frequency and high-frequency subbands of the image. The structure branch focuses on the low-frequency subband to restore overall structure, while the texture branch reconstructs intricate details using three high-frequency subbands. Key enhancements include the integration of an Efficient Transformer Block (ETB) in the structure branch for global information capture and a Residual in Residual Dense Block (RRDB) in the texture branch for detailed texture recovery. Additionally, we employ the CIEDE2000 color difference formula to ensure color consistency in the restored images. Our extensive testing demonstrates that WGDN outperforms existing methods in both qualitative and quantitative measures, establishing new benchmarks for the field of image inpainting.