高动态范围图像光照学习算法  

High dynamic range image illumination learning algorithm

作　　者：李彤 陈纯毅[1] 胡小娟[1] 于海洋[1] 李海兰[1] LI Tong;CHEN Chun-yi;HU Xiao-juan;YU Hai-yang;LI Hai-lan(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022

出　　处：《计算机工程与设计》2024年第9期2733-2741,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(U19A2063)。

摘　　要：为解决三维场景光照估计易受天气条件和拍摄视角等多方面因素影响的问题,提出一种基于深度学习的球谐系数估计方法,包括特征提取模块和光照估计模块。使用输入的图像对及预先计算好的光照参数训练网络,特征提取模块提取输入图像的特征,迭代改变球谐系数损失和渲染损失来优化网络,通过光照估计模块预测表达场景光照的球谐系数。实验通过与其它6种主流方法以及3种网络模型进行比较,所用评价指标均有显著性降低,实验结果表明,提出方法可有效恢复光照效果,定量定性地验证了恢复的光照在视觉上是真实的。To solve the problem that the illumination estimation of 3D scene is easily affected by many factors such as weather conditions and shooting angle,a spherical harmonic coefficient estimation method based on the deep learning was proposed,which included feature extraction module and illumination estimation module.The input image pairs and pre-calculated illumination parameters were used to train the network.The feature extraction module extracted the features of the input image,iteratively changed the spherical harmonic coefficient loss and rendering loss to optimize the network.The spherical harmonic coe-fficient of the scene illumination was predicted by the illumination estimation module.Compared with other 6 mainstream methods and 3 network models,the evaluation indexes used are significantly reduced.Experimental results show that the proposed method can effectively restore the illumination effect,and quantitatively and qualitatively verify that the restored illumination is real in vision.

关 键 词：深度学习 光照估计 球谐光照 虚实融合 光照一致性 高动态范围 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]