Research On Embroidery Style Simulation Based On CNN

With the continuous development of computer graphics,many scholars at home and abroad have studied the use of Non-Photorealistic Rendering to simulate the generation of oil painting,ink painting,random stitch embroidery,pyrography and other works of different artistic styles.Some scholars have also proposed embroidery simulation algorithms for embroidery works of art.Embroidery,with a long history,is a unique traditional art work in China.It is of great artistic value to weave all kinds of real scenery patterns on embroidered cloth by using needles and threads.Real embroidery art works are often used to decorate all kinds of daily necessities.For the simulation of embroidery art style,we can use the Non-Photorealistic Rendering to build the embroidery mathematical model and get the embroidery simulation result image.The existing embroidery simulation algorithm determines the direction of the needle and thread by analyzing the edge structure of the image,and then selects the type and color of the embroidery thread to simulate the target area to generate the needle and thread texture.However,the algorithm generally lacks the characteristics of flexibility and changeability.There are some problems in the obtained embroidery simulation results,such as the weak sense of needle and thread,and the single direction of needle and thread.Convolution neural network can extract the features of the target image through convolution calculation and match the features of the real embroidery artwork image.Finally,the needle and thread texture in the real embroidery artwork image are generated in the target image,and the artistic features of the embroidery simulation result image are similar to the real embroidery artwork.Compared with the existing embroidery style simulation algorithm,convolutional neural network has better generalization ability.In view of the shortcomings of existing embroidery style simulation algorithms,this paper proposes an algorithm based on convolution neural network for embroidery style simulation.Firstly,an embroidery simulation algorithm based on the convolution neuralnetwork of feature fusion is proposed.The network adopts the feature fusion structure to generate the result image with embroidery artistic style.The network model is composed of the generated network and the loss network.The generated network consists of VGG19 network encoder,matching module and decoder;the loss network uses VGG19 network to calculate style loss and content loss.In network training,aiming at optimizing the loss value calculated by the loss network,the parameters of the generated network are updated,and the steps of updating parameters are repeated until the specified number of times.After training,input the image into the trained network to get the embroidery simulation result image.Secondly,an embroidery simulation algorithm based on multi-scale two-channel convolution neural network is proposed.The network adopts multi-scale two-channel structure on the basis of feature fusion structure,and the network model is also composed of the generated network and loss network.The generated network is composed of multi-scale input,two-channel module and integration module.In order to make the result image not only have fine needle and thread texture,but also completely retain the detailed structure of the target image,the loss network uses Laplacian loss and the loss of VGG network composed of VGG19 network and VGG16 network.Finally,some of the original images are preprocessed to get the content image and embroidery art style image.In addition,the embroidery simulation result image obtained by convolution neural network is fused with mask image and background image to get the final embroidery simulation result image.The experimental results show that the embroidery style simulation algorithm based on convolution neural network proposed in this paper can show obvious texture features and various directions of stitches,and has good generalization ability.In addition,the embroidery simulation algorithm based on multi-scale two-channel convolution neural network can keep the details of the image more completely,and the needle and thread texture of the resulting image is clearer,more expressive and more close to the real embroidery works.