Novel Optimal Experimental Design Algorithm for Image Retrieval

Most of the existing optimal experimental design (OED) methods are based on either linear regression model or Laplacian regularized least square (LapRLS) model. This paper proposes a new active learning algorithm based on the second-order Hessian energy, which has the manifold learning capability. The algorithm selects those optimal samples which minimize the parameter covariance matrix of the Hessian regularized regression model, and overcomes the drawbacks of LapRLS. The experimental results on content-based image retrieval have demonstrated the effectiveness of the proposed approach.