Image Removal and Inpainting Based on FPGA and GAN

Abstract

Because irrelevant people are often taken in landscape, personal and group photo shoots, it is practical to find out how to remove them easily and effectively and restore the background. An improved PR-GAN is designed for image background inpainting by using semantic segmentation networks and using the idea of Generative Adversarial Network (GAN). Attention mechanism can improve PR-GAN to extract input features better and reduce the irrelevant factors and color discontinuity after removing characters because of the network depth. The improved PR-GAN runs on the platform of FPGA+CPU+GPU by building the FPGA hardware platform. The experimental results show that the PSNR and SSIM of the improved model reach 17.647 and 0.6320, respectively, in multi-person scenes.