A City Remote Sensing Image Positioning Algorithm Based On Semantic Information

Image positioning of cities is of great value in areas such as autonomous satellite navigation,urban real-world navigation,and urban 3D reconstruction.Traditional image image positioning and registration are based on feature point matching.Due to the large size,complex information,and other properties of the image,and the need for positioning of images of different sources and different time phases,the image is characterized There are many problems in point extraction and matching,such as large amount of calculation and many mismatches.In order to solve the above problems,this paper proposes a two-step repositioning image positioning algorithm based on basic information based on the characteristics of city prediction images.The first step of the algorithm is to pre-position the template image,and the second step is to perform accurate registration between the template image and the area obtained from the predetermined position.When pre-positioning the image,the algorithm first performs semantic segmentation based on deep learning on the urban remote sensing image,and obtains the road texture map and the feature distribution information of the urban remote sensing image.Based on the road texture map of the city,the similarity between the template image to be positioned and the sub-region in the source image is measured.Sort the obtained image similarities from high to low to obtain a queue of regions of interest,and the content of the queue is the pre-positioning information of the image.When performing precise registration between images,the template image is sequentially registered with the region of the region of interest queue obtained in the pre-positioning based on feature points.If a sufficient number of matching points can be obtained,and the homography matrix of the image can be obtained based on the matching,the algorithm flow is terminated,and the obtained homography matrix is used to correct the pre-positioned area to obtain accurate positioning.The advantages of the algorithm in this article mainly include the following two points:(1)A remote sensing image semantic segmentation network is designed based on the UNet framework.The network uses Res Net to replace UNet's original encoder to extract deeper image features;at the same time,a perforated convolutional layer is added between the encoder and the decoder to expand the network's receptive field.The network has achieved a good segmentation effect,and the average accuracy is about 2%higher than the original UNet framework.(2)Through the pre-positioning operation based on semantic information,the extraction of feature points is restricted to the range of interest,which reduces the influence of the detail information of irrelevant areas in the image on the feature point matching,thereby improving the feature point extraction and matching effectiveness.