[论文] Learning the Target Priors Before Image Translation: A Decoupled Train...

论文概要

研究领域: CV
作者: Keyan Hu, Mingtao Wang, Ziyu Zhou, Tiandong Shi, Haifeng Li, Ji Qi, Chao Tao
发布时间: 2026-08-28
arXiv: 2608.28517

中文摘要

遥感中的跨模态图像翻译必须保留源观测内容同时匹配目标域分布。现有方法从稀缺配对数据中联合学习目标先验和跨模态依赖性，忽视了一个关键不对称性：只有后者本质上需要跨模态对应。我们通过条件分数和去噪风险分析形式化了这种区别，并提出了LTP-BIT（Learning the Target Priors Before Image Translation），一种先验优先范式，将两个学习任务解耦。LTP-BIT首先从大规模未配对图像中学习目标域生成先验，然后保留预训练骨干权重并通过P-DART学习源条件控制，P-DART是一种参数高效的双流架构。控制实验表明，先验匹配和缩放主要改善目标域真实性，而实例保真度更强烈地依赖于条件适应。LTP-BIT在仅使用9.81%任务特定参数的情况下，在SAR-to-RGB和NIR-to-RGB基准上实现了最先进的性能。在QXS-SAROPT上，它仅使用25%的配对样本就保留了接近全数据的实例保真度。

原文摘要

Cross-modal image translation in remote sensing must preserve source-observed content while matching the target-domain distribution. Existing methods jointly learn the target prior and cross-modal dependence from scarce paired data, overlooking a key asymmetry: only the latter intrinsically requires cross-modal correspondence. We formalize this distinction through conditional-score and denoising-risk analyses and propose Learning the Target Priors Before Image Translation (LTP-BIT), a prior-first paradigm that decouples the two learning tasks. LTP-BIT first learns a target-domain generative prior from large-scale unpaired imagery, then retains the pretrained backbone weights and learns source-conditioned control through P-DART, a parameter-efficient dual-stream architecture. Controlled exp...

自动采集于 2026-09-01

#论文 #arXiv #CV #小凯