Open AccessRemote Sensing Image Building Extraction Method Based on Deep Learning
Author(s) -
Minshui Wang,
Mingchang Wang,
Guodong Yang,
Ziwei Liu
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1631/1/012010
Subject(s) - computer science , artificial intelligence , pixel , convolutional neural network , feature extraction , pattern recognition (psychology) , extraction (chemistry) , image (mathematics) , computer vision , function (biology) , chemistry , chromatography , evolutionary biology , biology
Using random patches and deeplabV3+ network can effectively improve the building extraction accuracy and ensure the integrity of building. First, acquisiting the image of a 5000×5000 pixel one, and using the random Patch Extraction Datastore function to create a number of random patches with the size of 224×224 pixels as network input images. Second, creating a convolutional neural network based on resnet50 by using the deeplabv3plusLayers function, and then projecting the learned discrimination features with lower resolution to the pixel space with higher resolution, to realise the automatic extraction of the building. Third, two images were input to verify the extraction accuracy of the trained network. The results showed that the Pixel accuracy of image 1 and image 2 reached 97.98% and 92.59%. Compared with other building extraction algorithms, this method has higher extraction accuracy. This method has strong expansibility and It can be used for automatic extraction of other feature types.