基于卷积神经网络的风格迁移泳装图案设计

Design of style transfer swimsuit patterns based on convolutional neural network

将深度学习中的风格迁移技术运用到服装图案设计领域正逐渐成为热门话题,但传统风格迁移图案设计的方法存在风格单一、纹理简单的缺点。本文提出一种辅助泳装设计师进行图案设计的方法,通过优化风格迁移中格拉姆矩阵结构,分别学习提取多张绘画作品和照片作为风格图生成创新的泳装图案设计图,同时运用Canny边缘检测算法和OpenCV图像处理库将作品从复杂背景中识别并分割出来,最终实现多风格图像输入的风格迁移辅助泳装图案设计方法。经过对比实验证明,本文生成的图案比其他风格迁移方法在风格满意度等方面获得更高评价。

With the penetration of artificial intelligence technology into various fields, the combination of artificial intelligence and art design provides a broader prospect for the intelligent design of clothing. As an active topic of deep learning of artificial intelligence, style transfer starts to be used in the fields of clothing pattern design and art painting. At present, there are many technical shortcomings in using style transfer technology for clothing pattern design. When style transfer based on the convolutional neural network is applied to clothing pattern design, the problems of monotonous color, simple texture and inability to remove redundant backgrounds arise. Therefore, this study explores an integration of Gram matrix and Canny edge detector to solve the problem of multi-style fusion and background segmentation in style transfer.In this study, in order to realize multi-style transfer, we first input multiple style images into the VGG-19 model, so that the layers designated as style output can extract the features of each style image and output them separately. We calculate the Gram matrix of each image separately, and weight all the obtained Gram matrices to form a new matrix. Therefore, the co-occurrence and correlation of each channel in the new matrix can represent the fusion style. In order to deal with the redundant backgrounds generated in the style transfer process and the non-rendered areas due to the features of swimsuit styles, we adopt the Canny edge detector algorithm and the OpenCV image processing library to perform operations such as rendering segmentation of images using the HSV interval differences of different rendered areas of swimsuit, and finally obtain the swimsuit pattern design drawings. Compared with other convolutional neural networks whose style transfer can only extract the style of one image, we optimize the structure of the Gram matrix, and can extract the style of multiple images to transfer at the same time. In the processing of the image generated by style transfer, by analyzing the structure and design features of swimsuit, a clothing image segmentation model applicable to swimsuits is established. In order to verify the effectiveness of this study in the field of clothing pattern design, we compare the effect drawings generated in this study with those generated by the style transfer method of other convolutional neural networks using the three metrics of questionnaire score, PSNR, and SSIM, and the results show that the method of this study obtains higher evaluation in all the three metrics.This study, by combining the painting art style with swimsuit pattern design, is able to design a large number of swimsuit patterns with multi-style fusion features at a very low cost and has great application prospects. There is still room for improvement in the accuracy of swimsuit image segmentation in this study, and further research will be conducted in this area.