基于注意力机制与三元组损失的少样本字体生成方法

Few-Shot Font Generation Method Based on Attention Mechanism and Triplet Loss

摘要: 针对跨语言少样本字体生成任务中风格特征表达不足和字形结构迁移不够精准的问题，本文提出一种融合注意力机制与三元组损失约束的少样本字体生成方法。该方法在对风格字体图像进行特征提取的基础上，引入上下文增强注意力机制，通过结合卷积操作与上下文感知特征变换模块，实现对字体笔画细粒度局部特征与整体结构信息的联合建模。随后，解码器对融合特征进行重建，生成目标字体图像。此外，本文引入三元组损失函数，从内容一致性与风格判别性两个维度对生成结果进行约束，以进一步增强字形结构一致性和风格表达能力。实验结果表明，该方法在少样本及跨语言字体生成任务中能够有效提升字体的结构保真度和风格迁移质量。

Abstract: To address the problems of insufficient style feature representation and inaccurate glyph structure transfer in cross-lingual few-shot font generation tasks, this paper proposes a few-shot font generation method that integrates an attention mechanism with triplet loss constraints. Based on feature extraction from style font images, a context-enhanced attention mechanism is introduced, which combines convolutional operations with a context-aware feature transformation module to jointly model fine-grained local stroke features and global structural information of glyphs. Subsequently, the decoder reconstructs the fused features to generate target font images. In addition, a triplet loss function is incorporated to constrain the generated results from both content consistency and style discriminability perspectives, thereby enhancing glyph structural consistency and style expression capability. Experimental results demonstrate that the proposed method effectively improves structural fidelity and style transfer quality in few-shot and cross-lingual font generation tasks.