MICPL: Motion-Inspired Cross-Pattern Learning for Small-Object Detection in Satellite Videos.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Traditional small-object detection methods rely heavily on visual patterns, offering limited feature learning capabilities.
• Existing schemes often utilize a single vision pattern, neglecting the potential of latent motion patterns for improved detection.
• Human perception effectively utilizes multi-pattern signals, including motion, for efficient small-object recognition.

Purpose of the Study:

• To address the limitations in small-object detection by exploring latent pattern learning.
• To propose a novel approach that captures motion patterns for moving small-object scenarios.
• To enhance feature learning by integrating both visual and motion information.

Main Methods:

• Proposed a motion-inspired cross-pattern learning (MICPL) scheme.
• Developed motion pattern mining (MPM) to extract motion patterns from time-dependent representations.
• Implemented motion-vision adaptation to correlate motion patterns with visual semantics, exploring cross-pattern interactions.

Main Results:

• Incorporating motion patterns with even simple detectors significantly improved state-of-the-art (SOTA) results in moving small-object detection.
• The MICPL scheme demonstrated adaptivity and advantages across two small-object-related tasks.
• Experimental validation confirmed the effectiveness of the proposed cross-pattern feature learning approach.

Conclusions:

• The integration of motion patterns offers a powerful new direction for improving small-object detection.
• The proposed MICPL scheme effectively captures and utilizes latent motion patterns for enhanced detection performance.
• This cross-pattern learning approach shows significant promise for advancing the field of small-object detection.