Research On The Style Transfer Algorithm Guided By Limited Quasi-Newton Method

With the development of deep learning technology in the field of artificial intelligence,convolutional neural network has achieved great success in many fields.Convolutional neural network not only has a lot of achievements in speech recognition,image recognition,image segmentation,natural language processing,but also has a continuous improvement in style transfer technology.In the aspect of image processing,the interest of image style transfer technology is to render a picture into a new painting with artistic characteristics under the condition of constant content.Image style transfer mainly refers to the technology of using machine learning algorithm to learn the style of artistic representative paintings and transfer this style to a given picture.Through the design and implementation of the currently studied artistic style algorithm based on deep learning,some problems are found,such as: the image distortion and poor accuracy in the implementation of image style migration by convolutional neural network,and the stylistic results are not clearly distinguished when the style migration is transmitted to the text.This paper proposes the following solutions to these problems:(1)In order to solve the problem of decreasing accuracy and image distortion of convolutional neural network in image style transfer,an image style transfer algorithm based on convolutional neural network is proposed.Firstly,the traditional texture reconstruction algorithm is analyzed,and the LBFGS optimization method,one of the quasi Newton methods is used to improve it.Then,the texture,color and visual information in the image are calculated by using the Gram matrix,and two high-level abstract feature representations of an ordinary image and a representative artistic image are extracted to generate a synthetic image with original content and artistic style.Based on the deep learning Keras framework with Tensorflow as the back end,an image style migration algorithm based on convolutional neural network is designed.(2)In order to solve the problem of poor accuracy of convolutional neural network in the process of text style transfer,this paper uses the Gram matrix to calculate the characteristics of text,and calculates the total style loss by calculating the difference between the Gram matrix of the style picture and the result picture,and takes the average value by combining the data of multiple feature layers,so as to improve the loss degree by using the L-BFGS optimization method which is one of the quasi Newton methods Fine tuning and adding image enhancement technology makes the image color more gorgeous,thus generating a synthetic text image withtext and artistic style.The algorithm proposed in this paper successfully solves the complexity problem and improves the accuracy at the same time.In the aspect of image style transfer,a proper number of iterations can be selected to observe the matching degree of the synthesized image.In the aspect of text style transfer,two style images can be transferred and the intensity of random noise can be adjusted.The proposed algorithm has been well applied and verified.