A Lightweight Detection Method for Remote Sensing Images and Its Energy-Efficient Accelerator on Edge Devices

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Embedded Systems

Background:

• Convolutional Neural Networks (CNNs) are vital for remote sensing image detection but often too complex for edge devices.
• Power-sensitive and resource-constrained remote sensing edge devices require efficient network solutions.

Purpose of the Study:

• To propose a lightweight CNN for remote sensing edge devices.
• To develop an energy-efficient CNN accelerator using Field-Programmable Gate Arrays (FPGAs).

Main Methods:

• Network weight reduction and optimization techniques were applied to minimize size and hardware deployment complexity.
• A reconfigurable and efficient convolutional processing engine was developed for the FPGA accelerator.
• Hardware optimization was performed specifically for the proposed lightweight network structure.

Main Results:

• The proposed network achieved higher accuracy with a smaller size.
• The CNN accelerator demonstrated a throughput of 29.53 GOPS and consumed only 2.98 W.
• The solution utilized 113 Digital Signal Processors (DSPs) and showed 1.1-2.5 times increased DSP efficiency.

Conclusions:

• The developed lightweight CNN and FPGA accelerator offer a superior solution for remote sensing edge devices.
• The proposed system is highly suitable for deployment on power-sensitive and resource-constrained remote sensing edge applications.