Orientation Feature Recognition Of Polarization Field In Cloudy Sky Based On Deep Learning

The sky polarization directional navigation has the advantages of no error accumulation and no electromagnetic interference.The symmetry of the solar meridian in the polarization azimuth diagram is very significant in light cloudy weather.Many scholars have developed the polarization navigation method to determine the course Angle by identifying the solar meridian.However when there are thick clouds in the sky,the polarization degree of this region is reduced and the polarization azimuth is abnormal due to the repeated scattering and absorption of sunlight by a large number of water droplets or ice crystal particles in the clouds.At this time,if the pixels with abnormal polarization azimuth in the target space with more clouds are directly fitted to extract the solar meridian features,then the wrong solar azimuth reference information in the carrier coordinate system is obtained.At present,the existing imaging polarization orientation methods have limited application scope in the weather with large area of thick cloud distribution.Therefore,this paper proposes a cloud segmentation method using improved UNET deep convolution neural network combined with polarization degree threshold,which can remove the anomalous pixel points corresponding to the cloud area in the polarization azimuth map,and obtain the solar direction based on Rayleigh scattering theory and solar vector constraint to achieve the polarization orientation of the sky.Firstly,the polarization model of cloudless sky was constructed based on Rayleigh scattering theory,and the actual sky polarization field map(polarization degree and polarization Angle map)and sky intensity map were obtained based on Stokes vector formula.The polarization Angle map was processed to obtain the polarization azimuth map.The polarization distribution of cloud region under cloudy weather is analyzed by using the theory of Mie scattering.It is found that the polarization degree in this region is too low,which leads to abnormal polarization azimuth.Then,an improved U-NET deep convolutional neural network for sky cloud segmentation is constructed.Respath module containing three residual structures is adopted to better integrate the coding area and decoding area of U-NET neural network.The binary segmentation image predicted by the neural network and the binary segmentation image with the polarization degree threshold of 0.1 were performed or calculated to obtain a binary segmentation image that could better segment the cloud and the blue sky.The binary image was used as a mask template to remove the abnormal pixel points in the polarization azimuth image.Secondly,the imaging sky polarization orientation method based on Rayleigh scattering theory and solar vector constraint is described in detail.The polarization orientation experiment is carried out by simulating clear weather and cloudy weather.An imaging sky polarization detection system is established.In order to solve the problem of large radial distortion of fisheeeye lens,a calibration method of zenith Angle of fishee-eye lens based on beacon light-emitting diode was proposed.Finally,the experiments of obtaining the solar azimuth in the carrier coordinate system are carried out in the actual cloudless weather and cloudy weather.Experiments on cloudless weather show that the proposed method has the smallest root mean square error(RMSE)of0.28° compared with the two methods(least square method and box separation method)for fitting the solar meridian.Experiments on cloudy weather show that the RMSE error of solar azimuth in carrier coordinate system obtained by this method is 0.42°.It can be seen from the experimental results,the method presented in this paper has good feasibility and robustness in the practical directional application of sky polarization navigation.