A Method Of Texture Synthesis Based On Convolutional Neural Network

Texture synthesis,as a significant branch of computer graphics,has always been the focus of scientific researchers.Texture synthesis technology has great application value not only in the field of image editing and restoration,but also in the field of large-scale scene generation and image rendering.Convolutional Neural Network is proposed based on visual cognitive mechanism.It has the characteristics of sparse connection and weight sharing,as well as the local translation invariance brought by the idea of spatial down-sampling.In recent years,it has made outstanding achievements in many fields such as computer vision,speech recognition and so on.Texture synthesis based on Convolution Neural Network has become one of the hotspots of current research.This paper mainly studies the texture synthesis method based on Convolution Neural Network.This paper focuses on the main methods of texture feature extraction technology,involving the traditional methods which are based on gray level co-occurrence matrix and Gabor filter for texture feature extraction,as well as the more novel method which is based on Clem matrix for texture feature expression.Based on the above work,an improved VGG-19 network model is proposed to improve the speed and quality of texture image synthesis.The main contents of this paper are as follows:(1)In order to improve problems of long training time and disappearance of gradient caused by excessive network layers,an improved method is proposed.Firstly,the VGG-19 network model of Convolution Neural Network is selected,and then the Batch Normalization layer is added after each convolution layer to speed up the network training efficiency.The experimental results show that the training speed of the improved Convolutional Neural Network proposed in this paper has been significantly improved.(2)Texture image synthesis based on improved Convolution Neural Network.In this paper,the texture synthesis model only uses the convolution layers and the pooling layers in the VGG-19 network model,but does not use the full connection layers and the classification layer.Firstly,the source texture image and a white noise initialized image are input into the Convolutional Neural Network respectively.After the forward propagation process,the texture features of the image are extracted,and the loss function is constructed.Then the network is trained to update the synthetic image.The experimental results show that the method used in this paper has obvious advantages over other texture synthesis methods in both subjective and objective evaluation,and the quality of image synthesis has been greatly improved.