Pedestrian Image Inpainting Based On GAN Network

In recent years,with the rapid development and research of computer vision and deep learning in academia,the advantages of deep learning methods in image restoration,feature extraction,image translation,and image generation have become increasingly prominent.Deep learning has increasingly become the research direction of researchers.Deep learning methods have shown promising results for extremely challenging restoration tasks,and these methods can generate image structures and textures that look reasonable.Since Ian Goodfellow and others proposed GAN(Generative Adversarial Network)in 2014.Generative Adversarial Network(GAN)has now become a research hotspot in the fields of image restoration,image translation,and image generation.This article focuses on the repainting of pedestrian images.Nowadays,surveillance cameras are widely deployed in every corner of the city.Therefore,a large number of pedestrian images can be obtained every second.How to automatically analyze and understand the content behind it has become an urgent research topic,which has obvious theoretical and application value.Obviously,the quality of the source image will seriously affect the subsequent stage of understanding.Therefore,this article will discuss how to recover damaged pedestrian images.Since surveillance cameras are mostly distributed outdoors,the collected images will inevitably show some tonal distortion.We believe that if not handled properly,tonal distortion may reduce the repair effect.Therefore,this article proposes a comprehensive network structure that can restore damaged pedestrian images while correcting tonal distortion.In order to achieve robust recovery,this paper adopts a network framework that generates confrontation.The existing GAN method is very common.And many GANs use multiple discriminators.But their multiple discriminators are used to achieve the consistency of images from the global and local.Their purposes and usages are consistent.At this point our approach is different.Our overall generator completes the translation of the image.And our global discriminator is used to correct the tone.For the inpainting part,we consider adding an additional local discriminator to refine the effect of the inpainting.In this way,we construct two generative adversarial losses,for two completely different functions,achieving the perfect combination of tone correction and image inpainting.Our method emphasizes the fusion of image inpainting and tone correction through the GAN network.