Implementation of Lightweight Convolutional Neural Networks with an Early Exit Mechanism Utilizing 40 nm CMOS Process...

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Disaster Management

Background:

• Unmanned aerial vehicles (UAVs) are crucial for early disaster detection, but processing their data efficiently remains a challenge.
• Current machine learning solutions for UAV data analysis face high computational costs and energy demands.
• Cloud computing lacks real-time processing capabilities, while edge computing struggles with efficiency for disaster relief.

Purpose of the Study:

• To develop an energy-efficient and computationally optimized convolutional neural network (CNN) model for fire detection using UAVs.
• To address the limitations of cloud and edge computing in real-time disaster monitoring applications.
• To enhance resource utilization in UAV-based disaster response systems.

Main Methods:

• A novel convolutional neural network (CNN) model with an early exit mechanism was designed for fire detection.
• The CNN model was implemented using TSMC 40 nm CMOS technology for hardware acceleration.
• Power-gating techniques were employed to reduce energy consumption by deactivating idle memory components.

Main Results:

• The implemented CNN model achieved a maximum accuracy of 81.49% in hardware.
• The CNN circuit completed fire detection in approximately 230,000 cycles with a modest parameter count (11.2 k).
• The hardware accelerator operated at 300 MHz, consuming 117 mW of power, demonstrating significant energy efficiency.

Conclusions:

• The proposed early-exit CNN model offers an effective and efficient solution for real-time fire detection in UAVs.
• Hardware acceleration and power-gating techniques significantly improve computational and energy efficiency for edge AI applications.
• This approach enhances the viability of UAVs for timely and resource-efficient disaster monitoring and response.