Research On Semantic Control Image Synthesis Technology Based On Deep Learning

With the rapid development of digital image technology and artificial intelligence technology,the combination of digital image with high-dimensional abstract features and artificial intelligence algorithm has become the development trend of computer image technology.In the conventional image processing technology,image processing and synthesis specific operations can be realized by standardized calculation methods.For the spatial information extraction of objects in the image,the traditional image algorithm also needs to cooperate with the 3D model to achieve through cumbersome manual operations.Compared with traditional digital image processing technology,the use of convolutional neural network algorithm will extract more abstract features from digital images.Combining the spatial structure features and environmental features of objects in the image,the spatial transformation of 3D objects in the image can be realized efficiently and intelligently.Therefore,the research on semantic control image synthesis technology based on deep learning has important theoretical significance and practical value.The research content of this paper comes from the project of"Teleoperation Software for Scientific Experiments".Taking objects in machine vision as research objects,aiming at images.Key technologies such as reconstruction and perspective reconstruction,foreground segmentation and background reconstruction,and semantic control image synthesis are studied.The main research work is as follows:Firstly,a study of common 3D generation network models is carried out in combination with voxel model data sets.According to the image space feature extraction problem,the network structure of 3D-CNN,3D-GAN and VAE model is designed,and the model corresponding training and test data set are established based on the input form of the network model.For the classification recognition and object reconstruction problem,the voxel mesh model is designed to preprocess the data set,and the structural performance evaluation of various network models is realized based on the training and test results of three commonly used 3D generation network models.It paves the way for the subsequent establishment of image reconstruction and perspective reconstruction network model,foreground segmentation and background reconstruction network model.Secondly,combined with the research of structural characteristics of 3D generation network model and image reconstruction and perspective reconstruction algorithm,the research of multi-parallel multi-cascade coding-decoding network model is carried out.Based on the established image reconstruction and perspective reconstruction network model,Blender is used to render training and testing image data sets of multi-class 3D chairs.Aiming at the spatial transformation problem of image foreground objects with semantic control,adjusting network parameters and training parameters to optimize the network model,and comparing with common coding-decoding network,capsule network and traditional algorithm,from image synthesis effect and block matching histogram algorithm to achieve qualitative and quantitative evaluation of the network model.Then,based on the established image reconstruction and perspective reconstruction network model,the research of foreground segmentation and background reconstruction algorithm is carried out.Combined with the structural characteristics of the 3D generation network model,a two-level joint coding-decoding network model is proposed.For the multitasking network model training problem,the image data set is enhanced by adding background information to the rendered image.Design multi-task combination loss function to realize multi-task joint training of network model.For the image synthesis problem of semantic control,the F-measure algorithm is used to qualitatively and quantitatively evaluate the network model and the traditional algorithm from the perspective of segmentation tasks,and the image reconstruction and perspective reconstruction network model are combined with the foreground segmentation and background reconstruction network model which implements the construction of a semantic control image synthesis network model.Finally,an experimental study is started on the spatial transformation of image objects combined with rotational semantic information.Develop real chair image data collection standards,and build a semantic control image synthesis experiment verification software platform.Design experimental protocols and conduct experiments with software platforms and data collection.The experimental results are compared and analyzed to verify the validity and practicability of the semantic control image synthesis network model in real environment.