Research And Implementation Of UAV Remote Sensing Image Stitching Based On Convolutional Neural Networ

Unmanned aerial vehicle(UAV)remote sensing stitching technology has been widely used in fields such as geospatial mapping and environmental reconnaissance due to its flexibility and timeliness.It can acquire a panoramic image,which provides the complete information required for the target scene and overcomes the limited field of view of a single image caused by the focal length and flight altitude of the UAV.This paper focuses on the research of UAV remote sensing image stitching technology,and its specific research contents and innovative work are as follows:Firstly,an UAV remote sensing image registration method based on a Siamese convolutional neural network is proposed to address problems of insufficient feature extraction and feature matching errors due to the complex background information of UAV remote sensing image and scale change of images to be registered.To a start,we design a feature extraction network based on the VGG-16 to effectively extract features of UAV remote sensing image.Next,the feature matching map of image pair to be stitched is obtained by using Siamese feature extraction network to get feature matching information.Finally,group convolution is introduced to enable the network estimate efficiently the transformation relationship between the images to be stitched.Moreover,an UAV remote sensing image registration dataset is constructed with the transformation relationship between image pair as the true label.The network is accuratly trained by composite loss function and self-constructed datasets.Experimental results demonstrate that our method achieves better results in terms of SSIM,MI,and RMSE metrics compared with other image registration algorithms.Secondly,the traditional gradual image fusion algorithm is improved to eliminate the stitching artifacts of UAV remote sensing images and obtain a better stitching effect.A Gaussian model is introduced in the calculation of pixel fusion weight to make the fusion process smoother.At the same time,a grayscale value threshold constraint is set for the region to control the pixel fusion rules,taking the differences between the pixels to be fused into account.It makes the fused image more natural.Finally,comparative analysis of experiments proves our methond can effectively improve the image stitching effect and achieve traceless stitching of UAV remote sensing images.Finally,a software platform for UAV remote sensing image stitching is developed using Py Qt,which can conveniently and friendly display the stitching process and effect of UAV remote sensing images.