In order to solve the problem of annotation scarcity in qualitative remote sensing, meta-learning has been used to semantic segmentation to achieve the best performance. Model-Agnostic Meta-Learning (MAML) can be compatible with any gradient descent-trained model and performs well on classification tasks with fewer samples. In this paper, the MAML and fully convolutional neural networks are fused which be trained and tested by RGB remote sensing images. After fusion with MAML, the network structure is divided into inner loop and outer loop. The increase of the outer loop is the increase of the generalization of the network, which has a good effect on the complex and changeable remote sensing images. Practically, SegNet and U-Net are combined with MAML, respectively. For building detection, the fused network test accuracy can be improved by 4.33% and 4.5% respectively. And for road extraction, the test accuracy is improved by 4.47% and 7.24% under the same experimental conditions, respectively. It proves that the MAML algorithm has a good performance on semantic segmentation to reduce training samples.
Few-Shot Semantic Segmentation for Building Detection and Road Extraction Based on Remote Sensing Imagery Using Model-Agnostic Meta-Learning
Lect. Notes Electrical Eng.
2021-10-30
11 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
Road Recognition From Remote Sensing Imagery Using Incremental Learning
Online Contents | 2017
|