基于全卷积神经网络的图像修复算法研究

北京理工大学珠海学院2020届本科生毕业论文Research on image inpainting based on Fully ConvolutionalNetworksAbstractImage inpainting is a difficult problem in the field of deep learning.Traditional imageinpainting algorithms are mostly based on image diffusion or patch,and usually requiresmoothness prior.This algorithms often perform poorly in scenes with large missing areasor strong semantic scenes.The development of deep learning technology has made imagerepair have a new research direction.Image inpainting al gorithms based on deep learningcan be applied to a wider range of scenes,greatly improving the repair effect.However,the depth generation model still has great room for improvement,such as single networkstructure,unstable training process,poor interpretability,and lack oftheoretical certificationfor super parameter selection.This paper bring forward an image inpainting algorithm based on FCN.With reference to theidea of GAN,the network structure of the context encoder is improved,so that the partiallymissing image is visually consistent with the global image after completion.In order totrain the consistency of the image completion network,the discriminant network is dividedinto two parts:the global discriminator takes the complete image as the input to identify theglobal consistency of the scene;The local discriminator only observes the quarter size areaof the original image centered on the mask to ensure the local consistency of the generatedimage.In addition,the loss function,training steps and other aspects are also improvedin this paper.With CelebA as the training data set,the semantic repair task that cannot beachieved by traditional algorithm is realized in the test phase.Keywords:Image inpainting Deep leaming Fully Convolutional Networks Genera-tive Adversarial Network