基于生成对抗网络的街景图像修复模型

A Streetview Image Inpainting Method Based on Semantic Prior Guidance

摘要: 针对现有街景图像修复方法难以有效处理复杂遮挡和保持语义一致性的问题，本文提出了一种基于生成对抗网络(GAN)的图像修复模型，旨在利用语义信息提升修复效果。该模型能够有效地提取和处理不同类别的语义信息，并将其融入图像修复过程，利用语义信息引导生成器生成更逼真、更符合语义的修复结果。在街景数据集上的实验结果表明，相较于传统方法，本文提出的模型能够生成更加自然逼真的修复结果，可以精准地还原车辆、道路、建筑物等物体的轮廓、纹理细节和颜色，有效消除图像缺失部分带来的视觉突兀感，显著提升图像的整体质量。未来可进一步拓展该模型的应用范围，将其应用于人脸图像以及其他类型场景的图像修复任务中，以期提升图像修复模型的泛化能力和修复效果。

Abstract: This paper proposes a novel generative adversarial network (GAN)-based image inpainting model for street view images, aiming to leverage semantic information to enhance inpainting quality. The model effectively extracts and processes different categories of semantic information, integrating them into the inpainting process to guide the generator in producing more realistic and semantically consistent results. Experimental results on street view datasets demonstrate that, compared to traditional methods, our proposed model generates more natural and visually appealing inpainted images. It accurately restores the contours, textures, and colors of objects such as vehicles, roads, and buildings, effectively eliminating the visual abruptness caused by missing parts and significantly improving the overall image quality. In the future, we plan to extend the application of this model to face images and other types of scenes to improve the generalization ability and inpainting performance of image inpainting models, enabling them to better address diverse image inpainting needs across various scenarios.