Fast and Accurate Object Detection in Remote Sensing Images Based on Lightweight Deep Neural Network

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Object detection in remote sensing images is challenging due to complex scenes, dense targets, and scale variations.
• Balancing model complexity and accuracy is crucial for real-world deployment.

Purpose of the Study:

• To propose a lightweight object detector for high-speed and high-accuracy detection in remote sensing images.
• To address the trade-off between model complexity and accuracy in object detection algorithms.

Main Methods:

• Developed a lightweight YOLO-like object detector incorporating efficient channel attention layers.
• Utilized differential evolution to optimize anchor configurations for scale variations.
• Evaluated the model on the RSOD and DIOR datasets.

Main Results:

• The proposed network achieved 5.13% and 3.58% higher accuracy than state-of-the-art lightweight models on RSOD and DIOR datasets, respectively.
• Achieved a detection speed of 58 FPS with under 10W power consumption on an NVIDIA Jetson Xavier NX.
• Demonstrated suitability for low-cost, low-power remote sensing applications.

Conclusions:

• The proposed lightweight object detector offers a superior balance of speed and accuracy for remote sensing.
• The model's efficiency and performance make it highly suitable for practical, resource-constrained remote sensing scenarios.