Research Of Real-Time Image Style Transfer Algorithm Based On Cross-granularity Learning

Image style transfer is a computer vision technology that renders and synthesizes original images into images of different styles.In recent years,with the rise of deep learning and generative adversarial networks,image style transfer has gradually become a research hotspot.Existing image style transfer methods only learn global image features and ignore the importance of local instance features in the image.In addition,there are limitations in diversity,efficiency and speed.This thesis proposes a real-time image style transfer algorithm based on cross-granularity learning to improve the existing methods.In order to ensure that the local instances in the image are more natural after style transfer,this thesis proposes an image style transfer algorithm based on cross-granularity learning.Specifically,the global encoder and decoder learn the style transfer of the global image,the instance encoder and decoder learn the style transfer of the instance image,and the cross-granularity learning learns the relationship between the global image and the features of the instance image through cross-granularity consistent loss.In this way,the model can benefit from both coarse-grained and fine-grained learning processes,and the quality of image style transfer is improved by incorporating instance information through joint training.In addition,this thesis integrates style labels and random style features into the training process,thereby improving the efficiency and diversity of style transfer.In order to reduce the complexity and inference time of the image style transfer model in this thesis,a lightweight algorithm for the real-time image style transfer model is proposed.Due to the special training mechanism and generator structure of the model in this thesis,the existing CNN based lightweight methods cannot be directly applied.In this thesis,the knowledge distillation algorithm based on residual feature matching is firstly used to transfer the knowledge of the pre-trained original model to the small model,and then the decoder based on the Super mechanism is used to automatically cut the network channel to reduce the cost of training.Finally,the neural architecture search is used to separate the training and searching of the model to find the optimal network architecture automatically,and only once training is required to obtain the lightweight model.In this thesis,a large number of comparative experiments are performed to evaluate the effectiveness of the image style transfer model based on cross-granularity learning and the lightweight model.The experimental results on two datasets show that the proposed algorithm has a better performance.