Research Of Image Compositing Technology Based On Deep Learning

Image compositing technology is a kind of computer technology that captures a specified target and composites it to a new image in the input image.Its research has certain practical significance.The traditional image compositing method is mainly based on the color information and prior knowledge of the image,which leads to limitations and can not meet the needs of complex applications.The deep learning method learns advanced semantic features by the convolutional neural network structure,which can help to overcome the defects and bottlenecks of traditional methods.This paper adopts deep learning strategy to study the problem of image compositing and provides an effective solution based on deep learning.This paper studies the key technologies of image compositing,and focuses on three aspects:foreground target matting strategy,background image completion algorithm,and foregroundbackground cloning strategy.Firstly,in the foreground matting strategy,a deep learning architecture with a combination of low resolution and high resolution network is established,which solves the problem of uneven quality of datasets and defines a reasonable network structure and loss function to ensure the quality and performance of deep learning model.Using the establishment of a hybrid dataset of natural images and synthetic images,the problem of lack of datasets is solved.The proposed image matting algorithm avoids the limitation of the condition of the input of the trimap for the study of traditional problems,achieves the automatic matting of foreground objects,and can obtain fine image matting results.Secondly,in the background completion algorithm,a two-stage image completion strategy is designed,including the semantic completion phase and the realism enhancement phase.In the semantic complement network,an effective content loss function is designed to ensure the consistency of the sample contribution.In the reality-enhanced network,the preliminary complementation result is used as an input to reduce the negative impact of the useless information in the input.Content loss and adversarial loss define a reasonable loss function to ensure that the final completion result is realistic.The function of complementing the missing area of any size and arbitrary shape is achieved,and the complementation result with the semantics and realistic details can be obtained.In addition,this paper explores the foreground and background image cloning strategy,proposes an unsupervised image cloning method,designs an image cloning strategy based on Lab color space,achieves the reservation of the foreground channel,and uses the deep learning model based on generator-discriminator structure,which can ensure the consistency of the light and shade of the cloning area,to predict the brightness channel of the image.A gradient loss function is designed to control the network training by using the composite loss function,which preserves the inherent semantic information of the foreground part after cloning and obtains a realistic cloning effect with the consistency of light and darkness.In a word,this paper explores the key technologies of foreground matting algorithm,background completion algorithm and foreground and background image cloning strategy,and provides a solution based on deep learning for image compositing technology research,which has certain reference significance and worth.