Remote sensing image road network detection based on channel attention mechanism.

Area of Science:

• Computer Vision
• Deep Learning
• Remote Sensing Image Analysis

Background:

• Road network extraction from remote sensing images faces challenges with high training accuracy but poor real-world results.
• Attention mechanisms in deep learning offer efficient ways to focus on relevant information and improve performance.

Purpose of the Study:

• To enhance road network detection in remote sensing images by integrating channel attention mechanisms with ResNet.
• To address the limitations of existing methods that show a gap between training accuracy and actual extraction performance.

Main Methods:

• Proposed SE-ResNet and ECA-ResNet models by combining channel attention mechanisms with the ResNet architecture.
• Applied these models to remote sensing image datasets, including Massachusetts roads (MR) and CHN6-CUG roads.
• Evaluated performance based on accuracy, loss convergence, and various detection metrics like precision, recall, IOU, and F1 score.

Main Results:

• SE-ResNet and ECA-ResNet showed comparable accuracy and convergence to LeNet7 and ResNet, with a slight increase in computational load.
• Despite similar convergence metrics, the proposed models achieved significantly better final road network detection results.
• The channel attention mechanism effectively guided the networks to prioritize road features and de-emphasize non-road features.

Conclusions:

• SE-ResNet and ECA-ResNet demonstrate significant improvements in the accuracy and quality of road network detection from remote sensing images.
• The channel attention mechanism is a valuable addition for enhancing deep learning models in this domain.
• ECA-ResNet, in particular, shows strong potential for practical applications in road network detection.