A Residual-Inception U-Net (RIU-Net) Approach and Comparisons with U-Shaped CNN and Transformer Models for Building...

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Building segmentation is vital for urban planning and map production.
• Current methods face challenges with global context modeling (CNNs) and high memory usage (Transformers).

Purpose of the Study:

• To evaluate and compare the performance of various Convolutional Neural Network (CNN) and Transformer models for building segmentation.
• To introduce and assess a novel architecture, Residual-Inception U-Net (RIU-Net).

Main Methods:

• Implemented and compared 10 models: U-Net variants (U-Net, Residual U-Net, Attention Residual U-Net, RIU-Net) with CNN encoders (Inception, Inception-ResNet, Xception, MobileNet), and Transformer models (Trans U-Net, Swin U-Net).
• Trained and evaluated models on the Massachusetts Buildings Dataset and Inria Aerial Image Labeling Dataset.

Main Results:

• RIU-Net achieved the highest IoU (0.6736), F1 score (0.7868), and test accuracy (92.23%) on the Inria dataset.
• Attention Residual U-Net and Trans U-Net showed strong performance on the Massachusetts datasets, with highest IoU, F1 scores, and test accuracy varying between models and dataset sizes.
• Residual U-Net also demonstrated competitive results on the Massachusetts datasets.

Conclusions:

• RIU-Net is a highly effective model for building segmentation, particularly on the Inria dataset.
• CNN-based U-Net variants (Residual U-Net, Attention Residual U-Net) and Trans U-Net offer robust performance for building segmentation on aerial imagery datasets.