A Lightweight Convolutional Neural Network Based on Visual Attention for SAR Image Target Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Remote Sensing

Background:

• Shallow Convolutional Neural Networks (CNNs) have limited feature extraction for Synthetic Aperture Radar (SAR) image target recognition.
• Improving recognition efficiency and handling imbalanced data in SAR images remain significant challenges.

Purpose of the Study:

• To design a lightweight CNN model for enhanced target recognition in SAR images.
• To improve feature extraction capabilities and address data imbalance issues.

Main Methods:

• Incorporation of channel and spatial attention bypass mechanisms to boost feature extraction.
• Utilization of depthwise separable convolutions to reduce computational cost and increase recognition efficiency.
• Introduction of a weighted distance measure loss function to mitigate the impact of data imbalance on minority class recognition.

Main Results:

• The proposed lightweight CNN model significantly reduces model size and iteration time compared to recent advanced networks.
• The model demonstrates comparable or improved recognition accuracy while effectively alleviating data imbalance effects.
• Experimental validation on MSTAR and OpenSARShip datasets confirms the model's efficacy.

Conclusions:

• The developed lightweight CNN offers a more efficient and accurate solution for SAR image target recognition.
• The attention mechanisms and weighted loss function effectively enhance performance, particularly in imbalanced datasets.
• This approach advances the field of deep learning applications in SAR image analysis.