TAI-SARNET: Deep Transferred Atrous-Inception CNN for Small Samples SAR ATR.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Synthetic Aperture Radar (SAR) target recognition faces challenges due to coherent speckle noise and high computational complexity in traditional deep learning models.
• Effective feature extraction and efficient processing are crucial for accurate SAR target identification.

Purpose of the Study:

• To develop an effective and lightweight Convolutional Neural Network (CNN) model for SAR target recognition.
• To address the limitations of traditional deep learning models in handling SAR data characteristics.
• To improve recognition performance, especially on small sample SAR datasets.

Main Methods:

• Proposed an Atrous-Inception module combining atrous convolution and inception module for rich global receptive fields and a lightweight architecture.
• Implemented a transfer learning strategy to leverage prior knowledge from optical and non-optical domains for SAR target recognition.
• Validated the model on the MSTAR dataset under standard operating conditions.

Main Results:

• Achieved a mainstream target recognition rate of 97.97% on ten types of MSTAR datasets.
• Demonstrated strong robustness and generalization performance on small, randomly sampled SAR target datasets.
• The proposed lightweight CNN model effectively mitigates speckle noise and reduces computational complexity.

Conclusions:

• The proposed lightweight CNN model with the Atrous-Inception module and transfer learning offers a superior solution for SAR target recognition.
• The method significantly enhances recognition accuracy and robustness, particularly for datasets with limited samples.
• This approach provides a computationally efficient and effective tool for SAR image analysis and target identification.