摘要
目的针对当前自动白平衡(AWB)方法在跨传感器应用中存在的泛化性差、数据采集成本高等问题,提出一种无需目标设备图像的跨设备AWB策略。方法构建融合动态色温校准与语义特征建模的方法。在校正阶段,选取多个典型色温下的白点构建分段对角映射矩阵,用于将训练图像重构至目标传感器响应域,实现数据分布对齐;随后,在训练阶段引入语义特征提取模块,以增强模型对跨设备光源估计的鲁棒性。结果实验在多个典型AWB数据集中均取得了领先的估计精度,具备良好的泛化能力和稳定性。结论所提方法实现了无需目标图像参与的跨设备AWB估计,大幅降低了数据采集成本,并显著提升了模型的部署效率。
To solve the problems of poor generalization and high data acquisition costs in cross-sensor applications of the current Auto White Balance(AWB)method,the work aims to propose a cross-device AWB strategy that requires no images from the target device.A method combining dynamic calibration at multiple typical color temperatures with semantic feature modeling was established.During the calibration,white points at multiple typical color temperatures were selected to construct a piecewise diagonal mapping matrix.This matrix was used to reconstruct training images into the target sensor's response domain,achieving data distribution alignment.Subsequently,in the training phase,a semantic feature extraction module was introduced to enhance the model's robustness for cross-device illuminant estimation.Experiments achieved excellent estimation accuracy across multiple standard AWB datasets,demonstrating strong generalization capability and stability.The proposed method enables cross-device AWB estimation without requiring any target device images,substantially reducing data acquisition costs and significantly improving model deployment efficiency.
作者
杨晟炜
方恩印
岳书威
YANG Shengwei;FANG Enyin;YUE Shuwei(Shanghai Publishing and Printing College,Shanghai 200093,China;Shenzhen Polytechnic University,Guangdong Shenzhen 518000,China)
出处
《包装工程》
北大核心
2025年第21期217-223,共7页
Packaging Engineering
基金
上海市人才项目(JS2024055)
上海市东方英才教师计划。
关键词
白平衡
颜色恒常性
颜色校正
域适应
white balance
color constancy
color calibration
domain adaptation
