Improving GAN based Image Inpainting by Object Classifier

Abstract

We present an improved inpainting method for the object image, by applying image classifier to the inpainting network. Conventional GAN based image inpainters worked well on scene images or human faces, but they also showed limitation with object images. In this paper, we propose an inpainting system for object images in which image classifier and image inpainter are combined. Our inpainting system has multiple inpainting networks which are fine-tuned for inpainting single object image, and those networks are applied to masked input image with the classification result of image classifier. With our system, we obtained improved inpainting result for the bicycle, bus, car, truck images and two percent improvement in the scoring method using Inception-ResNet-v2.