Enhanced Semantic Segmentation of Aerial images with Spatial Smoothness Using CRF Model

This paper addresses a deep learning method for high-resolution semantic segmentation in aerial images. U-net architecture promises end-to-end learning from basic ideas, making hand feature design deserted. However, the problem is gradually collecting information over larger image regions, making separating donations from different pixels. To solve this problem, the proposed training strategy is based on U-net, which contains two parts: contraction and expansion to segment foreground and background pixels. In addition, the significance of conditional random field (CRF) is applied to improve the accuracy of semantic segmentation. The proposed algorithm was evaluated on the Semantic segmentation of aerial imagery (Satellite images of Dubai) dataset, containing six common resources Land, Building, Road, Vegetation, Water, Unlabeled. The experimental findings reveal that the suggested approach outperforms other algorithms by achieving 0.99 accuracies and loss function 0.58.