Abstract. Image inpainting has always been a major problem in the field of scientific research. How to fill in the damaged area of the image to make the image look realistic is the main goal of the image inpainting task. In recent years, with the rapid development of machine learning, the researchers started using machine learning to assist in image inpainting. The most representative one is image inpainting based on generative models. The classic idea of restoration is to generate a restoration result with the highest quality and the most realistic appearance. However, as long as the result is reasonable, image inpainting allows for multiple restoration results. For this reason, the field of pluralistic image completion was born. This paper introduces pluralistic image completion and reviews past research in this field. This paper divides pluralistic image completion methods into three categories: VAE-based, GAN-based, and Transformer-based, and gives examples of various representative methods and the latest research in this field. This paper also introduces the available datasets and evaluation metrics. A discussion is given based on the performance of these three categories of methods and the prospects for the development of the field of pluralistic image completion. This review can serve as a reference for researchers in the field of image inpainting. It provides a table of datasets that can be used to evaluate and compare the listed methods and looks forward to possible developments and applications in this field in the future.
Read full abstract