Remote Sensing Image Ship Matching Utilising Line Features for Resource-Limited Satellites

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Geospatial Analysis

Background:

• Existing remote sensing image matching methods often rely on local features like SIFT, which can lead to excessive keypoints on backgrounds, reducing focus on main objects and increasing computational load.
• This inefficiency limits the performance of image matching, especially on resource-constrained satellite platforms.

Purpose of the Study:

• To develop an efficient and accurate image matching method for remote sensing scenes, particularly for ship matching on resource-limited satellites.
• To improve the focus on salient object features and reduce computational overhead compared to traditional local feature-based methods.

Main Methods:

• A novel keypoint extraction strategy, line feature based keypoint detection (LFKD), is proposed to leverage line features for selecting and filtering keypoints.
• LFKD enhances features at object corners and edges, significantly reducing false matches and improving object-centric representation.
• An end-to-end matching process incorporating a new crop patching function is introduced to further reduce complexity.

Main Results:

• The proposed method achieves a matching accuracy of 0.972 with low memory usage (313 MB) and fast testing time (138 ms).
• Extensive experiments demonstrate that LFKD improves matching accuracy when combined with existing Convolutional Neural Network (CNN) models.
• The method offers approximately a 50% test speed boost and a 30% memory saving compared to state-of-the-art approaches on both created and public datasets.

Conclusions:

• The LFKD method provides a significant advancement in remote sensing image matching, offering a balance between accuracy and computational efficiency.
• This approach is well-suited for deployment on resource-limited satellite systems, enabling more effective object recognition and tracking.
• The proposed method demonstrates broad applicability, enhancing various CNN-based descriptor models for improved performance in remote sensing applications.