Real-time scene classification of unmanned aerial vehicles remote sensing image based on Modified GhostNet

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Unmanned Aerial Vehicles (UAVs) are crucial for autonomous remote sensing image classification.
• Deploying deep learning models on embedded systems for real-time analysis faces resource limitations.

Purpose of the Study:

• To develop a lightweight deep learning network for efficient UAV remote sensing image classification.
• To balance computational efficiency and classification accuracy for embedded applications.

Main Methods:

• A novel lightweight network, Modified GhostNet, was designed based on GhostNet.
• Network modifications include altering convolutional layers and replacing the fully connected layer with a fully convolutional layer.
• Performance was evaluated on UCMerced, AID, and NWPU-RESISC datasets.

Main Results:

• Modified GhostNet reduced Floating Point Operations (FLOPs) from 7.85 MFLOPs to 2.58 MFLOPs.
• Memory usage decreased from 16.40 MB to 5.70 MB, with an 18.86% improvement in prediction time.
• Average accuracy increased by 4.70% on AID and 3.39% on UCMerced datasets.

Conclusions:

• The Modified GhostNet offers improved performance for lightweight networks in remote sensing scene classification.
• This network effectively enables real-time ground scene monitoring using UAVs.