Boosting Sparsity Constrained Bi-Linear Model for Object Recognition

Although the bag-of-visual-words (BoW) representation has received wide application, it ignores the spatial information. To tackle this problem, we propose to use ‘components’ as the higher-level visual elements to represent images. Then we formulate object recognition into a bi-linear model along with sparsity constraints to indicate two progressive linear relationships among a given concept and the two-level visual elements of images, yielding a sparsity constrained bi-linear model (SBLM). To further enhance robustness and leverage the nonconvexity of SBLM, we learn a combination of SBLM in a boosting-like procedure. We also prove that the optimization over each subset (while keeping the other subset fixed) of weighted SBLM is convex, hence the global optimum for each subset can be found. Accordingly, the optimization procedure can keep on refining the two parameter sets with strictly reduced objective cost during each boosting round. Experimental results demonstrate the effectiveness of the proposed boosting SBLM method. * Corresponding author. Tel.:86-10-62632267; fax: 86-10-62551993 E-mail address: [email protected] Digital Object Indentifier 10.1109/MMUL.2011.2