关键词: 图像修复, 约束对抗网络, 图像分辨率, 注意力机制

Abstract: Current image inpainting methods typically employ a single-resolution image as input for feature extraction, which limits the utilization of diverse features from different resolutions, resulting in suboptimal restoration of texture details and even artifacts. To enhance the inpainting performance, this paper proposes a multi-input multi-output constrained adversarial network for image inpainting. A multi-input low-resolution image generation module is designed. This module employs an encoder-decoder structure to generate global structural information from multiple low-resolution input images. A residual attention feature fusion module is constructed to merge the global features of low-resolution images with the fine-grained features of high-resolution input image, aiding in the completion of the high-resolution image. A constrained multi-output multi-scale adversarial loss function is proposed. This function ensures the consistency of the generated images in terms of local details and high-level structures, thereby guaranteeing the quality and authenticity of the multi-output images produced by the discriminator. Comparative experiments are conducted on the CelebA-HQ and Paris Street View datasets. The experimental results demonstrate that the proposed method outperforms five representative single-input single-output image inpainting networks both subjectively and objectively. The multi-input multi-output constrained adversarial network effectively leverages the distinct features from different resolution input images for image inpainting, generating more credible local details and improving the overall image inpainting performance.

Key words: image inpainting, constrained adversarial network, image resolution, attention mechanism