The Process Analysis Method of SAR Target Recognition in Pre-Trained CNN Models

Area of Science:

• Computer Science
• Artificial Intelligence
• Remote Sensing

Background:

• Convolutional Neural Networks (CNNs) show promise in Synthetic Aperture Radar (SAR) target recognition due to automatic feature extraction and translation invariance.
• However, the 'black-box' nature of CNNs hinders understanding their decision-making processes.
• Analyzing these models is crucial for improving their reliability and interpretability in SAR applications.

Purpose of the Study:

• To analyze the internal workings of pre-trained CNN models for SAR target recognition.
• To elucidate the roles of convolution, activation functions, and fully connected layers in feature extraction and decision-making.
• To clarify the specific SAR image features that CNN models focus on during target recognition.

Main Methods:

• Utilized four classical CNN architectures (AlexNet, VGG16, GoogLeNet, ResNet-50) trained on the MSTAR dataset for ten-category SAR target recognition.
• Applied a process analysis method to examine pre-trained CNN models.
• Investigated the contribution of individual CNN components (convolution, activation, full connection) to the recognition task.

Main Results:

• The analysis clarified the SAR image target features that pre-trained CNN models attend to.
• Convolutional layers function as image filters, activation functions introduce non-linearity, and fully connected layers further refine features.
• The study demonstrated the adaptability of these classical CNN models to SAR imagery.

Conclusions:

• A novel paradigm for analyzing pre-trained CNN models in SAR target recognition is presented.
• The findings enhance the understanding of deep learning models' performance in SAR image analysis.
• This work contributes to building more interpretable and reliable SAR target recognition systems.