Lightweight faster R-CNN for object detection in optical remote sensing images

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Object detection is crucial for remote sensing applications like urban monitoring and disaster prediction.
• Faster R-CNN offers high performance but requires substantial computational resources and storage.
• Model compression techniques are essential to address these resource constraints.

Purpose of the Study:

• To develop a lightweight Faster R-CNN model for satellite imagery analysis.
• To minimize performance degradation during model compression.
• To reduce computational and storage demands for efficient object detection.

Main Methods:

• A novel bi-stage compression approach combining aware training and post-training compression.
• Aware training utilizes mixed-precision FP16 computation to accelerate training and optimize memory.
• Post-training compression involves unstructured weight pruning and dynamic quantization.

Main Results:

• Achieved an average 25.6% reduction in model size.
• Reduced the number of parameters by an average of 56.6%.
• Maintained high mean Average Precision (mAP) across datasets.

Conclusions:

• The proposed bi-stage compression effectively creates a lightweight Faster R-CNN for remote sensing.
• The method balances significant model compression with minimal loss in object detection accuracy.
• Enables more efficient deployment of object detection models in resource-constrained remote sensing scenarios.