Study And Implement Of Edge-missing Image Inpainting Algorithm

Since the AlexNet deep neural network has achieved the first prize in the image classification competition,the application of deep neural networks in image processing has become more and more widespread.Image inpainting patches the missing pixel according to the surrounding pixel information when there are some missing parts in the input image.Most of the traditional image patching techniques use mathematical and statistical methods to patch the missing parts,for example,according to the characteristics of the distribution of RGB values.Although these techniques can repair the missing image,there are still some blur in the repaired pixel.And using deep learning to repair images can better restore the image details.Compared with traditional image inpainting algorithms,deep learning pays more attention to the semantic part of the image.Deep learning uses many convolutional layers to extract the semantic information in the image,and use the semantic information to repair the missing parts.This paper studies the existing inpainting technology and the encoder-decoder neural network structure used in the technology,and finds that the existing image inpainting technology cannot fill the missing part well when the edges of the image are vacant.Existing inpainting technologies mainly focus on local,small-area missing parts,and when the missing part is about one-third in the image,the patched part will appear blur and the edges of the object are not clear.Therefore,this paper improves on the existing network structure and proposes a low-resolution inpainting network,which can better repair a large range of vacant images.The low-resolution inpainting network is mainly composed of two parts,one is a low-resolution image generation network,and the other is a resolution expansion network.The low-resolution image generation network is responsible for patching the vacant parts of the image and generating the low-resolution patched image.The resolution expansion network is responsible for extending low-resolution images to the original resolution.Finally,this paper uses the Places2 dataset to verify the output of the network.Compared with the existing image repair technology,the low-resolution repair neural network has better results.