StateNet: Deep State Learning for Robust Feature Matching of Remote Sensing Images

Area of Science:

• Computer Vision
• Remote Sensing
• Deep Learning

Background:

• Image correspondence is crucial for remote sensing (RS) tasks.
• Existing methods often rely on global constraints or handcrafted features, limiting flexibility.
• Challenges remain in achieving robust and accurate feature matching for both rigid and nonrigid transformations.

Purpose of the Study:

• To propose a flexible and general deep state learning network for rigid and nonrigid feature matching in remote sensing.
• To introduce a novel mechanism for transforming match states into latent canonical forms, reducing randomness.
• To develop a network that can be easily integrated into existing architectures.

Main Methods:

• Developed StateNet, a deep learning network employing alternating steps for feature recalibration and transformation update.
• Introduced two novel, differentiable operations: adaptive dual-aggregation convolution (ADAConv) and point rendering layer (PRL).
• Evaluated the network's performance on diverse real remote sensing image pairs.

Main Results:

• StateNet demonstrated superior performance in feature matching compared to state-of-the-art methods.
• The proposed ADAConv and PRL operations effectively recalibrate features and update transformations.
• Experiments confirmed the robustness and universality of the StateNet approach across various image datasets.

Conclusions:

• StateNet offers a significant advancement in remote sensing feature matching.
• The network's flexibility and integration capabilities reduce the cost of establishing reliable correspondences.
• This deep learning approach provides a powerful tool for a wide range of feature-based visual tasks in remote sensing.