光学学报, 2017, 37 (11): 1129001, 网络出版: 2018-09-07
基于深度神经网络的空间目标常用材质BRDF模型 下载: 1020次
BRDF Model for Commonly Used Materials of Space Targets Based on Deep Neural Network
散射 双向反射分布函数 神经网络 深度学习 空间目标材质 scattering bidirectional reflectance distribution function neural network deep learning materials of space target
摘要
由于双向反射分布函数( BRDF)经验模型与半经验模型对材质散射特性描述时存在局限性,导致其拟合结果与实测数据的误差较大。针对此问题,基于深度神经网络(DNN)构建了一种适用于具有不同散射特性空间目标材质的BRDF模型。建立的深度神经网络模型基于TensorFlow实现,包含4个隐含层,并采用AdaDelta梯度下降法进行优化,结合Dropout方法进行正则。随机抽取材质测量数据的一部分作为训练样本,最终得到BRDF与入射天顶角、反射天顶角以及观测方位角的映射关系模型。大量的实验结果表明,建立的深度神经网络模型具有良好的材质特性描述能力,且对于相同材质,模型的拟合误差小于经验模型。
Abstract
When the bidirectional reflectance distribution function (BRDF) empirical model and semi-empirical model describe the scattering characteristics of the material, the limitation of these models for the description of different scattering characteristics results in large errors between the fitting result and the measured data. To solve the problem, a BRDF model suitable for commonly used materials on space targets with different characteristics is constructed based on deep neural network (DNN). The DNN model, which contains four hidden layers, is based on TensorFlow implementation. It is optimized by AdaDelta gradient descent method, and combined with Dropout method for regularity. Part of the material measurement data is randomly selected as the training sample, and finally the mapping relationships between the BRDF and the angles of the incident zenith, the reflection zenith and the observation azimuth are obtained. A large number of experimental results show that the DNN model has good ability to describe the scattering characteristics of materials, and the fitting error of the DNN model is less than that of the empirical model for the same material.
刘程浩, 李智, 徐灿, 田琪琛. 基于深度神经网络的空间目标常用材质BRDF模型[J]. 光学学报, 2017, 37(11): 1129001. Chenghao Liu, Zhi Li, Can Xu, Qichen Tian. BRDF Model for Commonly Used Materials of Space Targets Based on Deep Neural Network[J]. Acta Optica Sinica, 2017, 37(11): 1129001.