Research On Face Image Inpainting Algorithm Based On Deep Learning And Generative Adversarial Network

Image inpainting technology has always been a research hotspot in the field of computer vision and image processing.Repair the missing area of face image by some technology has great application prospects and commercial value for medical beauty,investigation and handling,cultural relic protection and so on.The traditional face image inpainting algorithm mainly uses the neighborhood information of the missing area to repair the damaged image,but the repair effect of these algorithms is often fuzzy and rough.In recent years,deep learning has gradually played a huge role in the fields of image generation and image semantic restoration.The image inpainting algorithm based on deep learning can learn more advanced features of images and has better repair effects than traditional repair algorithm.And the generative adversarial though the mutual confrontation of generative network and discriminator network,it is possible to generate false images that are very like the true images.Therefore,this paper applies deep learning techniques and knowledge of generative adversarial network to repair face image.The original generative adversarial network is the basis of the image restoration model,and then improve from three aspects: network structure,loss function and evaluation index.The network structure of generator uses a cascaded generation model from coarse to fine,and adds an improved dense net block.The discriminator adopts a two-layer discriminant model that is locally and globally discriminated.The loss function is a combination of minimizing reconstruction loss and counter-loss.The subjective visual analysis and objectiveness are used in the evaluation method.Firstly,the CelebA data set is preprocessed to obtain the training image and the test image.Then the training image with the missing area is input into the generator network to obtain the preliminary repair map.After the discriminator network's confrontation training and the loss function are continuously optimized,the final completion result is obtained.Through subjective and objective test results shows that the repair task of facial image area is up to 60% on the face image.In the peak signal-to-noise ratio and structural similarity,this paper is improved by 0.332-8.893 dB and 0.002-0.356 respectively.Therefore,this paper can not only complete the information for the missing image,but also make the repaired area and the overall figure more harmonious,and the effect is better than other repair algorithms.