A Modification method based on U-Net for the distorted pseudo edge of aerial initial orthophoto

Abstract The images captured by UAV camera have serious non-perspective distortion, and the overlap rate of heading and side direction is high. Only about 30% area of the image is available. The characteristics of small image frame and small ratio of base to height also lead to of model connection.In addition, the accuracy of image segmentation based on feature extraction is far from enough.Although the classical convolutional neural network can achieve effective image segmentation and edge calculation, but the resolution is declining in the process of forward propagation, which makes it difficult to achieve accurate segmentation of building edge when only using the features of the last layer.The above problems are the main reasons for the poor accuracy and serious distortion of the final synthetic aerial orthophoto image.To solve this problem, this paper proposes a U-Net based method to calculate and correct the distorted pseudo edge of aerial orthophoto.The object of study is the initial Orthophoto Image which is not synthesized by aerial photography.Firstly, based on the idea of U-Net, a neural network model with excellent performance in the field of image segmentation, the symmetrical network structure is used to fuse the high-dimensional and low dimensional features of the depth network to restore the high fidelity real boundary.Secondly, before the true value output, for the distorted features, brown method is used to find the superposition constraint positions of ideal feature points and corrected feature points, calculate the pseudo edge between distorted and undistorted, extract and prune, and retain the undistorted true value region.Finally, nested overlay and constraint detection are performed by combining the DEM of satellite images with the combined aerial orthophoto results.In the research and test, the detection accuracy statistics of internal industry encryption points and field control points with different scale accuracy are adopted, and the total coverage area is 0.5km²,more than 4000 buiding target data sets.The results show that the DOM detection error of the new aerial composite image and satellite image is less than 3 m and 9 m, which shows that the edge calculation and correction of aerial orthophoto composite image based on U-Net is efficient and feasible.