Classification via Sparse Representation of Steerable Wavelet Frames on Grassmann Manifold: Application to Target...

Area of Science:

• Computer Science
• Signal Processing
• Machine Learning

Background:

• Automatic target recognition (ATR) remains challenging due to variations in operating conditions like depression angle, configuration, articulation, and occlusion.
• Existing ATR methods struggle to robustly handle these extended operating conditions, limiting their real-world applicability.

Purpose of the Study:

• To propose a new classification strategy for automatic target recognition that addresses challenges posed by diverse operating conditions.
• To develop a novel representation model based on steerable wavelet frames and Grassmann manifolds for improved target classification.

Main Methods:

• A new representation model is developed using steerable wavelet frames, conceptualized as elements on Grassmann manifolds.
• Grassmann manifolds are embedded into a reproducing Kernel Hilbert space (RKHS) to apply kernel sparse learning.
• An overcomplete dictionary is formed from training samples in RKHS, used to sparsely encode query samples via a designed Grassmann kernel function.

Main Results:

• The proposed representation model, utilizing directional components of the Riesz transform, offers a novel way to capture target features.
• A quantitative measure of similarity is established using the Grassmann metric for the proposed representation model.
• A global kernel function is generated through Grassmann kernels, enabling effective sparse representation and classification.

Conclusions:

• The developed strategy demonstrates significant advantages in target classification, particularly under challenging and varied operating conditions.
• The integration of Grassmann manifolds, RKHS, and kernel sparse learning provides a robust framework for advanced automatic target recognition.