An Automatic Building Extraction Method Fused with Geometry Perception and Image Understanding

Abstract

The invention discloses an automatic building extraction method that integrates geometric perception and image understanding. Firstly, remote sensing data is selected and preprocessed, and then the obtained multi-band remote sensing image and normalized digital surface model are used as input feature maps, which are input to Feature learning is performed based on the encoder side of the improved deep residual network; then the high-level semantic features are input into the multi-scale efficient perception module of the network, and the low-level semantic features are combined with the high-level learning output of the multi-scale efficient perception module of the building. Hierarchical semantic features are input to the network decoder to obtain a binary object classification map; finally, the binary object classification map output by the network decoder is input to the network loss function layer to "drive" the network to learn to fit the target task The optimal weight parameters of , and finally the final building classification map is output by the network output layer, and the end-to-end automatic building extraction is completed. The present invention can achieve the best effect in terms of building extraction accuracy and extraction efficiency.

Description

technical field [0001] The invention belongs to the application of deep learning technology in the field of remote sensing image intelligent interpretation, relates to an automatic building extraction method integrating geometric perception and image understanding, and in particular relates to a method for extracting buildings from multi-source remote sensing data (aerial remote sensing image, airborne laser A method for automatically extracting buildings from radar data). Background technique [0002] Remote sensing is a modern applied technology science developed with the advancement of modern science and technology in the 1960s. It acquires ground target data by applying various sensors (photographs, scanners, and lidars, etc.), and processes the data. and analysis to obtain useful information of the detected target, so as to realize the understanding and description of the target. According to the height of the platform, remote sensing is generally divided into aerospac...

AI Technical Summary

Problems solved by technology

However, the current network learning strategy that uses the digital surface model obtained from LiDAR data as a network-assisted or additional image feature to extract buildings not only lacks sufficient mining of the spatial geometric information of LiDAR data, but also greatly increases the calculation of the network. quantity and parameter quantity

In addition, the output classification results lack detailed information, which is the main problem faced by the fully convolutional neural network structure.

Images