Missing image regions can be restored using inpainting techniques which have numerous applications in computer vision such as image restoration and content replacement. This research presents an improved model of Image Inpainting with context-aware completion using Generative Adversarial Network (GAN). This research takes the approach of a two stage GAN model where the generator learns to make plausible pixel level details from geometric and contextual information, and a discriminator that differentiates between real and fake content to improve realism. In this scenario, the generator first does a coarse paint of the image and then enhances the details subsequently. To enhance spatial and semantic consistency, contextual loss functions are used during the inpainting model training. To furthermore increase the network’s effectiveness of generating realistic inpainting outcomes under extensive missing regions, the authors add perceptual loss, adversarial loss, and style loss to fuse texture and structural integrity. Many datasets used as benchmarks lead to the conclusion that the proposed GAN-based inpainting model is more efficient in achieving high visual and structural coherence in images as compared to other approaches.
Sarker et al. (Wed,) studied this question.