Classification on the monogenic scale space: application to target recognition in SAR image.

Area of Science:

• Computer Vision
• Signal Processing
• Remote Sensing

Background:

• Synthetic Aperture Radar (SAR) imagery presents unique challenges for target recognition due to its inherent characteristics.
• Existing classification methods may struggle with non-linearly separable SAR target data and varying operational conditions.

Purpose of the Study:

• To introduce a novel classification strategy for SAR target recognition using monogenic scale space.
• To enhance target recognition accuracy by exploiting multidimensional signal characteristics and advanced classification frameworks.

Main Methods:

• Exploiting monogenic signal theory for feature extraction, capturing spatial-temporal information in SAR images.
• Integrating monogenic features into a sparse representation-based classification (SRC) framework.
• Employing kernel combination for classification to address non-linearly separable data, fusing multiple monogenic signal components.

Main Results:

• Developed a monogenic feature extraction method involving downsampling, normalization, and concatenation across scales.
• Implemented score-level fusion for SRC and composite kernel learning for robust classification.
• Validated the method on the Moving and Stationary Target Acquiration and Recognition (MSTAR) database under diverse conditions (noise, angle variations, structural modifications).

Conclusions:

• The proposed monogenic scale space-based classification strategy significantly improves SAR target recognition accuracy.
• The method demonstrates superior performance compared to baseline algorithms, including linear SVM, kernel SVM, and various SRC variants.
• The approach is effective even under challenging, non-ideal operating conditions, highlighting its practical applicability.