Research On Preprocessing And Mosaic Technology Of Multispectral Image In UAV Low Altitude Remote Sensing

UAV low-altitude remote sensing is an important way of monitoring the growth and physiological conditions of crops.However,due to the limitations of drones and environments,the images acquired by multispectral cameras mounted on drones are allways small size and distorted.In this paper,geometric distortion correction,vignetting correction and radiation uniformity correction methods are proposed for the distortion of farmland remote sensing multispectral images.The paper also analyzes the factors affecting multi-spectral image mosaicing of farmland UAV remote sensing,and proposes corresponding image registration,fusion and multi-image splicing path and cumulative error control method.Specifically,the main research contents and results of this paper are as follows:1)Zhang Zhengyou's camera calibration method is used to correct the lens distortion of the image.After correcting,the image corner position in both x-direction and y-direction error are less than 0.070 pixels,the average error is only 0.044 pixels.The distortion correction effect is good,which helps to reduce image registration error caused by distortion.2)The Retinex theory is applied to the vignetting correction of remote sensing images.To solve the problems of unstable correction quality and being time-consuming in traditional function approximation method and being halo and gray and spectral data distortion in multi-scale Retinex algorithm,a multi-scale Retinex algorithm with spectral recovery is proposed.The method realizes multi-spectral image vignetting correction with low-spectral distortion by estimating the global brightness of remote sensing image and introducing spectral recovery factor.The method is compared with the function approximation method based on Gauss model and the multi-scale Retinex algorithm.The experimental results show that the proposed algorithm can achieve better vignetting correction performance without halo and gray phenomenon.As for image sharpness,contrast and spectral quality,the proposed algorithm is also superior to the other two algorithms.Moreover,the spectral index also show good robustness.3)In order to reduce brightness difference between remote sensing images,a multi-spectral image radiation uniformity correction method based on real-time illumination information is proposed.The relationship between illuminance and image DN value is analyzed experimentally,and the radiation uniformity linear correction model and nonlinear correction model are proposed and compared.The results show that the correction accuracy of the nonlinear correction model is better than that of the linear correction model.However,due to the low efficiency of nonlinear correction,a strategy for selecting linear or nonlinear models according to the intensity of environmental illumination changes is proposed.4)Aiming at the problem of low resolution,low contrast of farmland remote sensing multispectral images whitch leads to the small number and uneven distribution of image features,this paper proposes a SIFT feature detection method based on principal component image which is calculated from principal component analysis.The principal component image with the largest variance in multi-spectral image is the target of SIFT detection,which can effectively improve the number,uniformity and validity of detected feature points in farmland multi-spectral images,improve the robustness of feature point detection,and help to improve the stability of multi-spectral image stitching.In addition,this paper studies the threshold influence in nearest neighbor second nearest neighbor ratio method,and determines the optimal threshold value of 0.85.Four high frequency and low frequency rule combination effects in the Laplacian pyramid fusion algorithm are compared and the result shows the combination of high frequency take-up rule and the lowfrequency hat function weighted average rule perform best.5)In order to reduce the influence of the cumulative error caused by the UAV factor and the mosaic method error,a stitch method based on stitch-to-stitch witch automatically optimize the stitch path is proposed.In order to decrease the error caused by the spatial transformation model,this paper proposed a cumulative error control method whitch uses the external parameters of synchronous acquired high-resolution visible image to modify the model,which can effectively eliminate the accumulated error in the mosaicing process.As for the defect of method based on stitch-to-stitch,the phase correlation method is used to optimize the detection area of feature points in the image,which improves the efficiency and accuracy of feature point detection and matching.