Research On Super-resolution Performance Analysis And Method Based On Flame Light Field Image

The light field camera used in the field of flame temperature measurement can realize functions such as layered temperature measurement and 3D reconstruction,which is of great help to study the combustion principle and improve the combustion system.However,the light field camera is limited by the hardware structure,and the spatial resolution of light field images needs to be improved,and most of the existing light field super-resolution algorithms are aimed at non-transparent Lambertian body.The study of the light field super-resolution method for translucent flames will help to improve the reconstruction accuracy of the three-dimensional temperature field of flames based on light field imaging technology.In this thesis,the non-transparent light field super-resolution method is applied to the light field images of translucent flames,and the reconstruction performance is analyzed.Based on the analysis results,super-resolution method for flame light field images is proposed and verified by experiments.The main work is as follows:(1)Collect flame light field images and create flame light field datasets.The Lytro Illum light field camera was used to shoot three flames of alcohol,candle and mineral oil,and the light field images were obtained by parsing with the LFToolbox toolbox,which provided benchmark data for the performance analysis of flame light field super-resolution.Then,the dynamic flame light field is collected,and the flame light field data set is created through the steps of analysis,clipping and downsampling,which solves the problem of lack of corresponding data during network training.(2)Apply the non-transparent light field super-resolution method to the flame light field image to evaluate the reconstruction effect.In the experiment,classical and mainstream super-resolution methods such as sparse coding,graph regularization,linear subspace projection,light-field convolutional neural network,and bicubic interpolation were selected for super-resolution reconstruction,and compared with the non-transparency of the light field public data set EPFL light field for comparison.The results show that the reconstruction effects of the two are quite different.For the optimization-based method,the influence of flame features or superimposed projections should be considered,and for the learning-based method,the flame light field dataset should be used for training.(3)The graph regularization method based on flame edge perception is proposed.Under the framework of the plug-and-play Alternating Direction Method of Multipliers algorithm,the super-resolution reconstruction performed by the graph regularization method is used to restore the high-resolution information of the flame light field,and the flame area is detected by combining the RGB and HIS criteria,and it is combined along the horizontal and the vertical direction is converted into a one-dimensional vector respectively,and the edge-preserving filtering is realized by the recursive filtering method.The flame area is filtered by the recursive edge-preserving filtering algorithm of domain transformation.The experimental results show that the graph regularization method based on flame edge perception can better preserve the flame edge information,and has the best effect on the peak signal-to-noise ratio evaluation index,and the maximum view consistency is improved by 21.66%.(4)The flame light field super-resolution network of multi-feature fusion is further proposed.Two parallel feature extraction branches are designed to extract different levels of flame features with different sizes of convolution kernels.The features between different views are aligned through deformable convolution to avoid the parallax estimation step of display.Finally,the redundant information of flame light field is fully utilized from the perspective of feature propagation and feature fusion.The reconstruction results show that this method can maintain the details of the reconstructed flame view,and significantly improve the peak signal-to-noise ratio and view consistency.