计算机科学
约束(计算机辅助设计)
翻译(生物学)
理论计算机科学
算法
人工智能
自然语言处理
算法设计
机器翻译
域代数上的
特征(语言学)
约束满足
电子邮件
约束满足问题
作者
Zhuo Zhang,Entong Zhu,Zheng Wu,Tianyu Chen,S Q Fan,Kehua Guo
标识
DOI:10.1109/tmm.2026.3711932
摘要
Image-to-image translation aims to learn a mapping between two domains, facilitating tasks such as style transfer and domain adaptation. In the absence of paired samples, it remains challenging to simultaneously preserve both global semantic consistency and local detail fidelity. Existing methods typically rely on cycle consistency and adversarial losses to constrain the generated results. However, these approaches often fail to adequately address the relationship between global and local features, making it difficult to maintain both global semantic coherence and local detail accuracy in the output images. To overcome these challenges, we propose a novel global-local constraint for unpaired image-to-image translation (GLCT). The key idea is to ensure gradient transformation consistency between the source and target domains by applying a gradient smoothing loss in the generator’s latent space. This enhances the stability and quality of image translation in unsupervised settings. Building on this, we introduce a semantic cycle consistency constraint to reinforce the mapping consistency between the source and target domains, thereby ensuring global semantic coherence during cross-domain translation. Additionally, a geometric structure alignment constraint is employed to match the geometric features between the source and generated images. This further captures fine-grained details and ensures accurate local detail restoration and structural integrity in the generated images. Extensive experiments on multiple benchmark datasets demonstrate that the proposed method significantly enhances visual quality and achieves state-of-the-art performance across various evaluation metrics.
科研通智能强力驱动
Strongly Powered by AbleSci AI