Multi-stitch Embroidery Image Generation Via Residual Attention Network

As an ancient art form,the related work of embroidery image synthesis has been widely followed by the academic.Real embroidery image often boasts vibrant colors,intricate textures,and diverse stitch type.This makes embroidery synthesis a challenging task.However,existing traditional methods and those based on Generative Adversarial Networks(GANs)for generating embroidery images suffer from problems such as color shift,chaotic textures,and loss of original structure.Additionally,previous works based on GANs did not take into account the impact of stitch types on embroidery.Therefore,this paper mainly explores and studies the multi-stitch embroidery synthesis:(1)This paper proposes a network framework for embroidery image generation via residual attention network.Based on the idea of residual attention,three attention masks are designed,namely color attention mask,texture attention mask and source attention mask.Under the unpaired datasets,the network can generate embroidery color and texture images and fuse them.So as to avoid the problems of color cast,messy texture and even losing the structure of the input picture in the generated embroidery picture.(2)This paper further update the above network and proposes a network framework for multi-stitch embroidery image generation via residual attention network.By adding a multi-stitch module,the diversity of stitch styles in the results is greatly enhanced.The proposed multi-stitch module can match the appropriate stitch types according to the shape characteristics of color regions,making the results have various stitch styles and closer to the real embroidery.In the research process,a white filling technique was found,which greatly improved the stability of embroidery texture generated by network in the prediction stage.It solves the problem that the texture cannot be generated normally occasionally in the generation process.(3)A multi-stitch embroidery dataset was produced in this paper.This dataset consists of pairs of reference images and corresponding embroidery images,with each embroidery image being labeled with a type of embroidery stitch type.Currently,the dataset includes three stitch types: tatami stitch,flat stitch,and satin stitch.According to research,this dataset is not only the first publicly available embroidery dataset that is labeled with stitch types,but also the largest embroidery image dataset currently available for network learning.Extensive qualitative and quantitative experiments demonstrate that the embroidery images generated by the proposed network framework in this paper are superior to existing methods.In user study and comparative experiments,the color of results is closer to the input,and the embroidery texture is more realistic.And the texture contains three stitch styles.In quantitative experiments,the results in this paper has lower scores in Fréchet Inception Distance(FID)and Learned Perceptual Image Patch Similarity(LPIPS)than other methods,which shows that the embroidery results distribution is closer to the real embroidery image distribution.