Feature extraction by optimizing attributed scattering model toward sparsity

The existing algorithms for parameter estimation of attributed scattering model (ASM) usually encounters difficulties in selecting model order, improving estimation accuracy and decreasing computational burden. It is a fact that sparsity is being accepted as a desirable property of a model. The incremental sparse Bayesian learning (ISBL) method is therefore applied to realize the sparsity-driven continuous parameter estimation in this paper. Compared to the nonquadratic regularization algorithms with an overcomplete dictionary matrix, the ISBL method has advantages in continuous parameter optimization, hybrid model optimization and computation efficiency. Experimental results with the Air Force Research Laboratory (AFRL) “Backhoe Data Dome” demonstrate that the ASM optimized by the ISBL method can achieve desirable sparse solution for the problem of inversion scattering and accurately extract geometrical features.