Prompt-Driven Knowledge Distillation for Remote Sensing Object Detection.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Accurate object detection in remote sensing is crucial for analyzing complex, multi-scale targets.
• Existing knowledge distillation methods struggle with remote sensing data's unique challenges, including long-tail distributions and error propagation.

Purpose of the Study:

• To develop an advanced knowledge distillation framework (PDKD) tailored for remote sensing object detection.
• To enhance model adaptability, address data biases, and mitigate error propagation from teacher models.

Main Methods:

• Proposed a Prompt Driven Knowledge Distillation (PDKD) framework.
• Introduced Scale-Decoupled Feature Prompting (SDFP) for scale-specific knowledge transfer.
• Implemented Semantic Visual Co-Prompting (SVCP) using CLIP for long-tail category enhancement.
• Integrated a Self-Correcting Prompting (SCP) module to minimize error propagation.

Main Results:

• The PDKD framework achieved 49.0% mAP on the DOTA dataset with a 1x training schedule.
• Demonstrated improved performance in identifying multi-scale and multi-directional targets.
• Showcased enhanced adaptability and reduced error propagation compared to traditional methods.

Conclusions:

• The PDKD framework effectively addresses limitations of standard knowledge distillation in remote sensing.
• The proposed modules (SDFP, SVCP, SCP) contribute to superior object detection performance.
• This research offers a promising direction for accurate and efficient remote sensing object detection models.