Image Inpainting Based On Multiple Methods Of Deep Learning

In the process of daily acquisition and transmission of digital images,there are many reasons for their content to be interfered.With the rapid development of communication technology,people's life and image information begin to become inseparable,which also makes the solution to the problem of image loss increasingly sophisticated.Digital image inpainting is an important research direction in the field of computer vision.In recent years,with the excellent performance of deep learning in image processing,more and more research teams begin to use the methods related to deep learning to deal with image inpainting.In this paper,we use a variety of deep learning methods to study several key problems in image repair tasks.The main work consists of the following parts:(1)At the first,starting from the task of image style transfer,we collected and summarized the existing image style transfer methods at home and abroad,compares the traditional image style transfer methods based on non-parametric methods with the image style transfer methods based on deep learning methods,and show the convolutional neural network especially the depth of the multilayer convolution neural network applied in image feature extraction and advantage.In these methods,the applicable scenarios of different style transfer networks are introduced in detail and summarized as follows: The deep network will make the time of image generation result increase,the network parameter will be a lot;The training data of the pre-training network will interfere with the network's perception of the image content when it is used;It is difficult to control the result of the generated image when we use the Generative Adversarial Networks,because it needs high quality image data set and artificial hyperparameters.The main purpose of this part of work is to provide theoretical and experimental basis for the following research.(2)Using the structural advantages of U-Net network and the feature extraction capability of convolutional neural network,we propose an image inpainting method based on U-Net network structure.This method can make use the content of the image itself as the prior knowledge to repair the image without the need of pre-training.In the loss function definition of the network,Content similarity loss is used to replace the original Mean Square Error loss to ensure the quality of the generated image.Theresults show that the restoration image generated by this network is better than that of similar pre-training network.(3)Under the influence of the idea of separable convolutions,the original U-Net network is improved,some ordinary convolution in the submodule is modified to depthwise separable convolutions,and it is applied to the watermark removal task of digital image.The results show that the model does not reduce the quality of generated images while reducing the network parameters and computational complexity.It is proved that the depthwise separable convolutions is equivalent to the ordinary convolution,and provides a reliable experimental scheme for the research of watermark removal.