Promoting Connectivity of Network-Like Structures by Enforcing Region Separation

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Reconstructing network-like structures (roads, canals) from aerial images is crucial for mapping.
• Existing deep learning methods often struggle with connectivity and accuracy in these structures.
• Aerial image analysis requires robust methods for identifying linear features.

Purpose of the Study:

• To introduce a novel connectivity-oriented loss function for training deep convolutional networks.
• To improve the reconstruction of network-like structures, specifically roads and irrigation canals.
• To enhance the accuracy and connectivity of predicted structures in aerial imagery.

Main Methods:

• Proposed a connectivity-oriented loss function based on disconnections between background regions.
• Applied the loss function to train deep convolutional networks (convnets).
• Evaluated the loss function on small image crops to capture short segments.
• Tested on standard road benchmarks and a new dataset of irrigation canals.

Main Results:

• Convents trained with the proposed loss function significantly improve road connectivity.
• The method effectively closes gaps in predicted roads and canals.
• It also prevents false positive predictions by penalizing unwarranted disconnections.
• Skeletonizing the output of networks trained with this loss yields state-of-the-art maps.

Conclusions:

• The novel loss function enables highly accurate reconstruction of network-like structures from aerial images.
• This approach achieves state-of-the-art results in road and canal mapping.
• The loss function is versatile and can be integrated into existing deep learning training setups without modification.