Image Style Transformation Based On Linear Model

With the widespread application of Generative Adversarial Network(GAN)and transfer learning in the field of computer vision,great progress has been made in the direction of image generation,such as generation of restoration of image,super-resolution of image and image style transfer.Among them,the image style transfer can be widely used in the fields of photo synthesis,film synthesis,decoration design and privacy protection,and it can also be used in industrial networks and intelligent security.Image stylization is a form of artistic expression of style conversion,imitating the artist's creative techniques,and has a certain appreciation value.This paper uses the Ghost module in the deep neural network to generate more feature maps from cheap linear operations,and then uses the attention mechanism to transform various styles of images.The model is based on the latest research results of style transfer,that is,on the basis of the generative confrontation network of illustration style transfer,a more effective generative network method is used to optimize the original generative confrontation network.The Ghost module is introduced into the generation model of image style transfer,which generates more feature maps with less parameters.An attention mechanism is added to the generative network,and the interdependence between feature channels is modeled through machine learning,thereby improving performance and efficiency of network.We have created datasets of various styles including autumn style,winter style,stick figure style datasets,and conducted experiments on the datasets created and the currently widely cited public datasets.The results show that the improved method proposed in this paper will not reduce the image quality,and the generated image is natural,which conforms to the visual habits of the human eyes.In addition,it overcomes the problems of high computational cost and high computational resource consumption in the image style transfer model.In the subjective and objective evaluation indicators,the generated images have better evaluation results.