Design And Implementation Of Improved Image Inpainting Neural Network Based On DeepFill

In recent years,with the rapid development of software technology,computer visionbased image restoration technology has begun to play a role in various fields such as entertainment,medical,archeology,cultural relic restoration,and security.Image restoration has also received the attention of academia and has become a hot research topic in computer vision.At present,some traditional methods or deep learning-based methods have been used to inpaint images.In some cases,it can achieve a certain inpaint effect,but there are also many problems,such as bad inpaint result on irregular mask,can not be guided by the user to generate results,the texture is not detailed enough,and so on.In order to solve these problems,this paper improves the existing network and obtains a new neural network,which achieves a better inpainting result.The main research results are as follows:(1)Partial convolution and gated convolution are used to replace the ordinary convolution layer in the original generation network,so that the network that can only inpaint the regular rectangular mask can inpaint the irregular mask now.The concat layer is used to transfer the low-level image features of the low layer to the high layer,so that the inpaint result has more details.(2)WGAN-div is used to replace the original WGAN-GP as the improved solution of the main GANs,which achieves the goals of faster convergence and more complete mathematics,and solves problems such as mode collapse,gradient disappearance,and training difficulties better.(3)Add user guidance module.The edge of the image is extracted using the HED Edge Detection model as one of the input dimensions during training,simulating the user's hand-drawn input.At the same time,a user guidance module is added to improve the partial convolution,so that users can finally guide the network to generate their desired results by drawing lines.Finally,my network,the original network and the Partial Convolution network are compared horizontally,and a method based on double-blind testing is used to subjectively evaluate the result of image restoration of each model.