A Multi-Scale U-Shaped Convolution Auto-Encoder Based on Pyramid Pooling Module for Object Recognition in Synthetic...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Unsupervised representation learning (RL) struggles with limited labeled data in synthetic aperture radar (SAR) object classification.
• Existing methods often neglect discriminative details and distinctive SAR image characteristics, leading to performance degradation.

Purpose of the Study:

• To propose an unsupervised multi-scale convolution auto-encoder (MSCAE) for enhanced SAR object classification.
• To simultaneously extract global features and local characteristics of targets.
• To address performance deterioration in SAR classification due to limited labeled data.

Main Methods:

• Developed a U-shaped architecture with pyramid pooling modules (PPMs) for multi-scale feature extraction.
• Incorporated compact depth-wise separable convolution and deconvolution to reduce parameters.
• Integrated SAR speckle prior knowledge and a speckle suppression restriction into the objective function.
• Utilized structural similarity index metric (SSIM) for reconstruction loss, comparing with improved Lee sigma filtered images.

Main Results:

• The MSCAE effectively learns multi-scale features, capturing both global and local target characteristics.
• Experimental results on the MSTAR dataset demonstrated significant effectiveness in SAR object classification.
• The model showed robust performance under standard and extended operating conditions.

Conclusions:

• The proposed MSCAE model enhances SAR object classification by effectively learning discriminative features from limited labeled data.
• Simultaneous extraction of global and local information, coupled with speckle suppression, improves classification performance.
• The unsupervised approach offers a viable solution for SAR image analysis challenges.