非透明光场超分辨率方法在火焰上的性能分析

Performance analysis of the non-transparent light fieldsuper-resolution method on flames

目的:超分辨率是光场应用领域的一个研究热点,但现有算法大多针对非透明朗伯体,其在半透明火焰光场上的适应性和重建效果尚不明确,需要进一步评估。方法:拍摄蜡烛、酒精和矿物油三种火焰光场图像,选取稀疏编码、图正则化、线性子空间投影、卷积神经网络、双三次插值等经典和主流的超分辨率方法进行超分辨率重建,并与公共数据集EPFL的非透明光场进行比较。结果:结果显示,半透明火焰重建精度高于非透明光场,且在图正则化方法上好于卷积神经网络方法;内外焰叠加投影到多视角图上,视图一致性优于非透明光场;视觉感知和光场EPI与非透明光场基本一致。结论:针对火焰光场的超分辨率的研究,应结合火焰发光辐射模型与光场成像模型进行。

Aims:Super resolution is a research hotspot in the field of light field application.But most of the existing algorithms are aimed at the non transparent Lambertian body;and its adaptability and reconstruction effects in the light field of translucent flame are not clear.So it needs to be further evaluated.Methods:Three flame light field images of candles,alcohol and mineral oil were taken.Sparse coding,graph regularization,linear subspace projection,convolutional neural network,bicubic interpolation and other classic and mainstream super-resolution methods were selected for super-resolution reconstruction and were compared with the non-transparent light field of the public data set EPFL.Results:The results showed that the reconstruction accuracy of translucent flame was higher than that of non transparent light field;and the graph regularization method was better than the convolution neural network method.The inner and outer flames were superimposed and projected onto the multi-view image;and the view consistency was better than that of the non-transparent light field.Visual perception and light field EPI were basically the same as the non-transparent light field.Conclusions:For the super-resolution research of the flame light field,the flame luminescence radiation model and the light field imaging model should be combined.