A robust object detection approach using boosted anisotropic multiresolution analysis

In an unconstrained environment, adaptive classifiers produce improved recognition. Fast discrete curvelet transform has recently gained attention due to its ability to capture singularities along curves far away from smooth regions. Therefore, curvelet coefficients contain enhanced representation of image details at different scales and orientations. We propose a new approach for object class detection based on curvelet feature subspace obtained using Kernel PCA (KPCA) and learned using AdaBoost scheme [1]. The first contribution of current paper is a unique representation of an image called curvelet feature subspace that preserves global structure, and supports reliable detection of singularities along curves which play a considerably important role in recognition. Second contribution of our proposed method is an adaptive selection of features obtained using anisotropic style multiresolution analysis for robust object detection of varied inter-class, and intra-class attributes. Our proposed method achieved better detection rate compared to state-of-the-art schemes.