A Building Extraction Method for Remote Sensing Images Based on Dual-path Feature Pyramid Network
Although convolutional neural network has achieved great success in building extraction from remote sensing images in recent years,it still faces some problems,such as multi-scale target recognition,fuzzy target boundary segmentation,and improving the recognition accuracy of unbalanced targets.In order to solve the above problems,a dual path feature pyramid encoder-decoder structure for remote sensing image is proposed to build ex-traction in the paper.The high-resolution network is used as encoder and the feature pyramid with attention mecha-nism is used as decoder to improve the recognition ability of the same target at different scales.Boundary network is added into the encoder to enhance the learning of target boundary features and enhance the extraction accuracy of target boundary recognition.The cross-entropy and Dice loss are weighted to enhance the accuracy of unbalanced target extraction.Finally,the experiment are carried out on WHU aerial image and WHU satellite image Ⅱ to e-valuate the method,and the intersections over union reach 90.0%and 71.1%respectively.
building extractionfeature pyramidconvolutional neural networkhigh-resolution networkremote sensing images
万方数据
维普
