Research On Key Technologies Of Image Background Removal Based On Generative Adversarial Network

Image background removal is one of the hot topics in computer vision research in recent years.It is widely used in film special effects production(green screen background),daily photo processing(id photo changing background),wearable devices(virtual reality)and other scenes.It is an extension of image understanding and analysis.In recent years,deep learning has developed rapidly in image processing,and its excellent ability to fit nonlinear functions makes it play a very good role in image segmentation,and has made breakthroughs in succession.The existing background removal methods based on deep learning can extract features through neural network,which has good generalization,but has higher requirements for image annotation.Address these questions.Relevant research work is carried out in this thesis,and the following achievements are achieved:1.Based on partial public data set and automatic labeling method,a background removal data set is created,which is completed by filtering high-quality portrait images and synthesizing them to random background.2.A new pre-segmentation module based on improved Deeplab V3+ is proposed.Through multi-scale fusion in the pooling layer of spatial pyramid,the global information is guaranteed and the details are not lost.The feature information is extracted by Xception in the encoding stage,and jump connection is added in the decoding stage.Multi-scale information is extracted by combining the shallow and deep features of the image.The experimental results show that compared with the original network structure,it has a certain improvement effect.3.An image background removal method based on multi-scale discriminator is proposed.Based on the network structure of generative adversarial network and the idea of residual network,an encoder-decoder matting network is designed.By selecting an appropriate deep convolutional network,deeper features can be extracted.In order to extract features of different levels,multi-scale discriminant network is introduced to further optimize the segmentation effect.Two image segmentation method in this thesis,the research has been in a public data set experiment,achieved the desired purpose,comparing the current methods and automatic method based on the Trimap has quite beyond the results,the experimental results show that image segmentation method based on the generated against network on the precision of the background to remove the image has very good ascension.