Text Guided Image Inpainting Based on Generative Adversarial Network

Abstract

In recent years, the generative adversarial network (GAN) has achieved good results in the field of image inpainting. At the same time, the research on the text-generation image method shows that the text-generated image content has achieved good results, and text description can generate semantic content with visual relevance. Therefore, combining GAN and text information can be used to generate image information, which can both make up for the deficiency of GAN and provide guidance for image inpainting. In this paper, a text-guided image inpainting model based on generation adversarial network is proposed. This module contains spatial and channel information to focus on the information related to inpainting task. However, the original generation adversarial network has the problem of gradient disappearance and the quality of the generated image is unstable. In order to stabilize the training process and improve the image repair effect, the model uses the method of least squares generation to improve the network, and further stabilizes the repair effect by introducing the perceptual network.