Research On Image Inpainting Methods Based On Semantic Perception Deep Model

Image inpainting is a task of synthesizing missing content in an image based on known information,so that the observer cannot perceive that the image has been damaged.As an important content in computer vision,image inpainting has been widely concerned for its important application value in many fields,such as digital cultural heritage protection,picture and video re-editing,television and film post-production.Traditionally,according to the correlation between image pixels and the similarity of image content,researchers mainly use mathematical and physical methods to achieve inpainting.With the rapid development of deep learning techniques in academia and industry,especially the emergence of various generative networks,image inpainting based on deep models can obtain better results by extracting image semantic feature effectively.In which,the consistency of structure and the integrity of semantics are both key points.Besides,the generalization ability of networks is also important in image inpainting for irregular holes.To address the issues of distorted structure,ambiguous semantics and weak generalization ability in image inpainting,we propose two works as follows:1)We propose a semantic image inpainting method based on densely connected generative networks.Image inpainting is a research of inferring the true data distribution based on the known prior knowledge of an image.Therefore,extracting the latent space feature accurately for obtaining prior knowledge is important.We propose a new end-to-end framework for semantic image inpainting.Through by the densely skip connections across a battery of symmetric encoder-decoder groups,different levels of semantic features can be effectively combined to explore deeper image understanding.In order to fit the data distribution of real images and obtain more realistic inpainted results,we train our parameterized generative model based on WGAN-GP.Compared with several state-of-the-art algorithms on two benchmark datasets,the results show that our method is the best both on subjective and objective evaluation.2)We propose a semantic image inpainting method based on consistency regularization.Deep networks fit the data distribution by learning a linear or non-linear mapping relationship based on large samples.Therefore,the generalization ability is important for deep model.In the field of image inpainting for irregular holes,the randomness also requires that deep model should own good generalization ability.Based on these,we introduce the consistency regularization mechanism into a generative network for image inpainting.Specifically,we minimize the self-consistent loss between the output values of the homogeneous image data with different perturbations.3)We minimize the self-consistent loss between the latent features of homogeneous image data to promote the model’s ability of extracting semantic feature.Compared with the baseline methods and state-of-the-art methods,experimental results have verified the effectiveness of the proposed method.