基于投影流網絡和嵌入空間損失的紋樣風格遷移

摘 要： 針對編解碼器框架造成的圖像重建誤差、恢復偏差以及圖像風格遷移方法存在內容泄露和局部偽影等亟待解決的問題，提出了一種基于投影流網絡以及嵌入空間損失的紋樣風格遷移模型。通過融合投影流網絡與矢量量化變分自編碼器實現無偏風格遷移，并對風格特征進行精細編碼和匹配，保留更完整圖像內容細節，且捕獲關鍵風格特征；設計了一種計算風格化圖和風格圖的嵌入空間損失，融入總體損失函數以保證風格特征的均勻分布，減少風格差異；以云錦為例進行織物紋樣風格的創新設計。結果表明：基于投影流網絡和嵌入空間損失的紋樣風格遷移的遷移效果，在內容評價指標結構相似性（Structural similarity，SSIM）、內容損失上的分值較對比模型分別提高86.21%、54.29%、20%和32.58%、18.68%、18.99%；風格評價指標Gram損失為4.5×10^-6，較對比模型提高近一倍，表明該方法有效平衡了內容保留與風格遷移的需求，提高風格化效果。該模型在改善內容泄漏的同時避免了內容失真和風格過度覆蓋，有效捕捉復雜風格特征和色彩層次，促進紋樣創新設計，增加風格遷移在紡織服裝領域的可能性。

關鍵詞：風格遷移；嵌入空間損失；矢量量化變分自編碼器；內容泄漏；投影流；紋樣

中圖分類號：TS941.26

文獻標志碼：A

文章編號：1673-3851 （2025） 03-0237-09

引文格式：朱昱儒，侯玨，楊陽，等. 基于投影流網絡和嵌入空間損失的紋樣風格遷移［J］. 浙江理工大學學報（自然科學），2025，53（2）：237-245.

Reference Format： ZHU Yuru，HOU Jue，YANG Yang，et al. Pattern style transfer based on projection flow networks and embedding space loss［J］. Journal of Zhejiang Sci-Tech University，2025，53（2）：237-245.

Pattern style transfer based on projection flow networks and embedding space loss

ZHU Yuru^a， HOU Jue^a，b， YANG Yang^a，b， LIU Zheng^b，c

（a.School of Fashion Design amp; Engineering; b.Key Laboratory of Silk Cultural Heritage and Product Design Digital Technology， Ministry of Culture and Tourism; c.Zhejiang International Institute of Fashion Technology， Zhejiang Sci-Tech University， Hangzhou 311199， China）

Abstract：Aiming at the problems that need to be solved urgently， such as image reconstruction error， recovery biases caused by coded framework， and content leak and local artifacts during the style transfer process， a pattern style transfer model based on projection flow network and embedded space loss was proposed. Firstly， by integrating the projection flow network with the vector quantized-variational autoencoder， unbiased style transfer was realized， and style features were finely coded and matched to retain more complete image content details and capture key style features. Then， a method was designed to calculate the embedded space loss of stylized graph and style graph， and the overall loss function was incorporated to ensure the uniform distribution of style features and reduce style differences. With Yunjin brocade as an example， the innovative design of fabric pattern style was carried out. The results showed that the transfer effect of pattern style transfer based on projection flow network and embedded space loss scored higher by 86.21%， 54.29% and 20%， and 32.58%， 18.68% and 18.99%， respectively in terms of the content evaluation metrics of structural similarity （SSIM）， and content loss compared to the comparison models; with a Gram loss of 4.5×10^-6， the model nearly doubled the performance of comparative models， indicating that this method effectively balanced the needs of content retention and style transfer and improved the stylization effect. The model improves content leak while avoiding content distortion and style over-coverage， effectively captures complex style features and color hierarchies， facilitates innovative design of patterns， and increases the possibility of style transfer in the textile and apparel field.

Key words：style transfer; embedding space loss; vector quantized-variational autoencoder; content leak; projection flow; patterns

0 引 言

圖像風格遷移技術是將一幅圖像的風格特征應用到另一幅圖像上的技術，廣泛應用于服裝設計^［1-3］、紋樣創新^［4-6］等場景。……