Building detection with convolutional networks trained with transfer learning
Peer reviewed, Journal article
Published version
Åpne
Permanent lenke
https://hdl.handle.net/11250/2984224Utgivelsesdato
2021Metadata
Vis full innførselSamlinger
- Import fra CRIStin [3628]
- Institutt for byggfag [296]
Originalversjon
Šanca, S., Oštir, K., & Mangafić, A. (2021). Building detection with convolutional networks trained with transfer learning. Geodetski vestnik, 64(04), 559-593. 10.15292/geodetski-vestnik.2021.04.559-593Sammendrag
Building footprint detection based on orthophotos can be used to update the building cadastre. In recent years deep learning methods using convolutional neural networks have been increasingly used around the world. We present an example of automatic building classification using our datasets made of colour near-infrared orthophotos (NIRR-G) and colour orthophotos (R-G-B). Building detection using pretrained weights from two large scale datasets Microsoft Common Objects in Context (MS COCO) and ImageNet was performed and tested. We applied the Mask Region Convolutional Neural Network (Mask R-CNN) to detect the building footprints. The purpose of our research is to identify the applicability of pre-trained neural networks on the data of another colour space to build a classification model without re-learning. Rezultati klasifikacije stavb na ortofotu se uporabljajo kot vir za vzdrževanje katastra stavb. V zadnjih letih se za klasifikacijo stavb v svetu vse bolj uveljavljajo metode globokega učenja z uporabo konvolucijskih nevronskih mrež. V raziskavi predstavimo primer samodejne klasifikacije stavb z uporabo lastnih podatkovnih zbirk, izdelanih iz barvnih bližnje infrardečih ortofotov (BIR-R-G) in barvnih ortofotov (R-G-B). Preizkusili smo detekcijo stavb z uporabo predučenih uteži podatkovnih zbirk Microsoft Common Objects in Context (MS COCO) in ImageNet. Za detekcijo stavb smo uporabili Mask Region Convolutional Neural Network (Mask R-CNN). Namen raziskave je preizkusiti uporabniško vrednost globokega učenja za detekcijo stavb z uporabo predučenih uteži na podatkih drugega barvnega prostora s ciljem izgradnje klasifikacijskega modela brez ponovnega učenja.